{"people":[{"i":0,"slug":"andrej-karpathy","name":"Andrej Karpathy","topic":"foundations","why":"Builds GPT from an empty file, then explains every line. The best place to start.","yt":"@AndrejKarpathy","ytId":"UCXUPKJO5MZQN11PqgIvyuvQ","ytSubs":1760000,"x":"karpathy","av":"https://pbs.twimg.com/profile_images/1296667294148382721/9Pr6XrPB_200x200.jpg","mode":"all","nv":12,"np":80},{"i":1,"slug":"3blue1brown","name":"3Blue1Brown","topic":"math","why":"The clearest visual math on the internet, including the neural network and transformer series.","yt":"@3blue1brown","ytId":"UCYO_jab_esuFRV4b17AJtAw","ytSubs":8740000,"x":"3blue1brown","av":"https://pbs.twimg.com/profile_images/1024330921367859200/n5_VT4zH_200x200.jpg","mode":"all","nv":151,"np":70},{"i":2,"slug":"welch-labs","name":"Welch Labs","topic":"foundations","why":"Slow, beautiful explanations of why deep learning works.","yt":"@WelchLabs","ytId":"UConVfxXodg78Tzh5nNu85Ew","ytSubs":913000,"x":"welchlabs","av":"https://pbs.twimg.com/profile_images/2073904738102390784/pi0lrbLC_200x200.jpg","mode":"all","nv":78,"np":41},{"i":3,"slug":"statquest","name":"StatQuest","topic":"foundations","why":"Statistics and machine learning, one small idea at a time.","yt":"@statquest","ytId":"UCtYLUTtgS3k1Fg4y5tAhLbw","ytSubs":1720000,"x":"joshuastarmer","av":"https://pbs.twimg.com/profile_images/1410315179854811145/aOs7Ec0M_200x200.jpg","mode":"all","nv":190,"np":40},{"i":4,"slug":"umar-jamil","name":"Umar Jamil","topic":"foundations","why":"Long whiteboard builds of transformers, LLaMA, diffusion and RLHF from scratch.","yt":"@umarjamilai","ytId":"UCtAcpQcYerN8xxZJYTfWBMw","ytSubs":92600,"x":null,"av":"https://yt3.googleusercontent.com/prohEQLNtch3TRelWvLjr9aoUMBecNtLHqwO7WYlY0uM1yRP_69aa-gxINvsghoUyi2GvPJRv5w=s0","mode":"all","nv":25,"np":0},{"i":5,"slug":"sebastian-raschka","name":"Sebastian Raschka","topic":"foundations","why":"Author of Build a Large Language Model (From Scratch). Careful and practical.","yt":"@SebastianRaschka","ytId":"UC_CzsS7UTjcxJ-xXp1ftxtA","ytSubs":102000,"x":"rasbt","av":"https://pbs.twimg.com/profile_images/1661187442043486209/a3E4t1eV_200x200.jpg","mode":"all","nv":314,"np":80},{"i":6,"slug":"jeremy-howard","name":"Jeremy Howard","topic":"foundations","why":"fast.ai: practical deep learning taught top-down.","yt":"@howardjeremyp","ytId":"UCX7Y2qWriXpqocG97SFW2OQ","ytSubs":151000,"x":"jeremyphoward","av":"https://pbs.twimg.com/profile_images/1279600070145437696/eocLhSLu_200x200.jpg","mode":"ai","nv":165,"np":5},{"i":7,"slug":"stanford-online","name":"Stanford Online","topic":"foundations","why":"Full Stanford lectures: CS336 language models, CS25 transformers, CS229.","yt":"@stanfordonline","ytId":"UCBa5G_ESCn8Yd4vw5U-gIcg","ytSubs":1330000,"x":null,"av":"https://yt3.googleusercontent.com/UE4m1O9zmzlEViC33IbGBU0idM-rbGzN1NWag8xMt6JHBG5SmuWmJdF1_uQhhGKLjLiUXvOTwA=s0","mode":"ai","nv":447,"np":0},{"i":8,"slug":"alexander-amini","name":"Alexander Amini","topic":"foundations","why":"MIT 6.S191, Introduction to Deep Learning, refreshed every year.","yt":"@AAmini","ytId":"UCtslD4DGH6PKyG_1gFAX7sg","ytSubs":358000,"x":null,"av":"https://yt3.googleusercontent.com/ytc/AIdro_k517G0Dt8QgsF86GKzYkG3Rs8jSGCiYGJw04lx4_1KXQ=s0","mode":"all","nv":97,"np":0},{"i":9,"slug":"serrano-academy","name":"Serrano.Academy","topic":"foundations","why":"Friendly, visual intros to transformers, attention and embeddings.","yt":"@SerranoAcademy","ytId":"UCgBncpylJ1kiVaPyP-PZauQ","ytSubs":197000,"x":null,"av":"https://yt3.googleusercontent.com/ksK0A4M7Y7FYaASCJvOo7gcVYsx38FcK30aj7CTEwMjTlrPCgNPaj4S9Hvg2TEE5r_2dCXxHB9k=s0","mode":"all","nv":76,"np":0},{"i":10,"slug":"ai-coffee-break","name":"AI Coffee Break","topic":"research","why":"Letitia explains new papers in a few minutes, with drawings.","yt":"@AICoffeeBreak","ytId":"UCobqgqE4i5Kf7wrxRxhToQA","ytSubs":65400,"x":"AICoffeeBreak","av":"https://pbs.twimg.com/profile_images/1370853286727782408/Q2X-CT11_200x200.jpg","mode":"all","nv":113,"np":42},{"i":11,"slug":"yannic-kilcher","name":"Yannic Kilcher","topic":"research","why":"Long paper readings, line by line.","yt":"@YannicKilcher","ytId":"UCZHmQk67mSJgfCCTn7xBfew","ytSubs":333000,"x":"ykilcher","av":"https://pbs.twimg.com/profile_images/473191190477537280/bWjYb1Rb_200x200.jpeg","mode":"ai","nv":418,"np":66},{"i":12,"slug":"two-minute-papers","name":"Two Minute Papers","topic":"research","why":"Short tours of new graphics and AI research.","yt":"@TwoMinutePapers","ytId":"UCbfYPyITQ-7l4upoX8nvctg","ytSubs":1840000,"x":"twominutepapers","av":"https://pbs.twimg.com/profile_images/1218465731123335170/H68927B9_200x200.jpg","mode":"all","nv":1072,"np":27},{"i":13,"slug":"machine-learning-street-talk","name":"Machine Learning Street Talk","topic":"research","why":"Long technical interviews with researchers.","yt":"@MachineLearningStreetTalk","ytId":"UCMLtBahI5DMrt0NPvDSoIRQ","ytSubs":221000,"x":"MLStreetTalk","av":"https://pbs.twimg.com/profile_images/2045594158510010368/BSqtMnpz_200x200.jpg","mode":"ai","nv":255,"np":26},{"i":14,"slug":"dwarkesh-patel","name":"Dwarkesh Patel","topic":"research","why":"Deeply prepared interviews with the people building AI.","yt":"@DwarkeshPatel","ytId":"UCXl4i9dYBrFOabk0xGmbkRA","ytSubs":1460000,"x":"dwarkesh_sp","av":"https://pbs.twimg.com/profile_images/1925260306684813315/NjNQZmhZ_200x200.jpg","mode":"ai","nv":115,"np":49},{"i":15,"slug":"neel-nanda","name":"Neel Nanda","topic":"research","why":"Mechanistic interpretability, taught by one of its founders.","yt":"@neelnanda2469","ytId":"UCBMJ0D-omcRay8dh4QT0doQ","ytSubs":14800,"x":"NeelNanda5","av":"https://pbs.twimg.com/profile_images/1542528376845615104/hWSoUGaN_200x200.jpg","mode":"all","nv":48,"np":63},{"i":16,"slug":"artem-kirsanov","name":"Artem Kirsanov","topic":"research","why":"Computational neuroscience meets machine learning, beautifully animated.","yt":"@ArtemKirsanov","ytId":"UCR2uRTQ53V_egXKFflMMaaw","ytSubs":375000,"x":null,"av":"https://yt3.googleusercontent.com/-2Qk9xw4zLGEOAGLKs_aYk_ufdrllaOa0UNDpN2d4h_gXTnJHinucqtWOnyrrofpRmp-U8fncw=s0","mode":"all","nv":36,"np":0},{"i":17,"slug":"latent-space","name":"Latent Space","topic":"agents","why":"The AI engineer's podcast. swyx and Alessio talk to the people shipping.","yt":"@LatentSpacePod","ytId":"UCxBcwypKK-W3GHd_RZ9FZrQ","ytSubs":445000,"x":"swyx","av":"https://pbs.twimg.com/profile_images/2073162797354217472/hNny55eF_200x200.jpg","mode":"ai","nv":301,"np":14},{"i":18,"slug":"ai-engineer","name":"AI Engineer","topic":"agents","why":"Every talk from the AI Engineer conferences: agents, evals, harnesses.","yt":"@aiDotEngineer","ytId":"UCLKPca3kwwd-B59HNr-_lvA","ytSubs":1000000,"x":null,"av":"https://yt3.googleusercontent.com/ajVemEB89DAOemsbfuMY6ZOWXJAACx3cbty9z21jeqRKODaVkDBSRun1b1xfQJljEsziOWS_Mg=s0","mode":"ai","nv":1383,"np":0},{"i":19,"slug":"indydevdan","name":"IndyDevDan","topic":"agents","why":"Agentic coding workflows with Claude Code and friends.","yt":"@indydevdan","ytId":"UC_x36zCEGilGpB1m-V4gmjg","ytSubs":154000,"x":null,"av":"https://yt3.googleusercontent.com/BMf5YUX1T-_JNAeS6bDrxU1Dw1DGDI9AEUSE2IMb0v5hUD2I0k0wEedFw2hJeLsyPq2v75Q0TQ=s0","mode":"ai","nv":215,"np":0},{"i":20,"slug":"cole-medin","name":"Cole Medin","topic":"agents","why":"Building agents and local AI stacks, start to finish.","yt":"@ColeMedin","ytId":"UCMwVTLZIRRUyyVrkjDpn4pA","ytSubs":231000,"x":"cole_medin","av":"https://pbs.twimg.com/profile_images/1870529436770004992/JeHbC_cQ_200x200.jpg","mode":"ai","nv":219,"np":45},{"i":21,"slug":"theo","name":"Theo","topic":"agents","why":"Opinionated takes on models, harnesses and dev tools. Filtered to AI.","yt":"@t3dotgg","ytId":"UCbRP3c757lWg9M-U7TyEkXA","ytSubs":578000,"x":"theo","av":"https://pbs.twimg.com/profile_images/2106484435382333440/vXOVE8EX_200x200.jpg","mode":"ai","nv":916,"np":16},{"i":22,"slug":"mario-zechner","name":"Mario Zechner","topic":"agents","why":"Author of Pi, the minimal open coding-agent harness.","yt":null,"ytId":null,"ytSubs":null,"x":"badlogicgames","av":"https://pbs.twimg.com/profile_images/1553485821767991296/87k3l720_200x200.jpg","mode":"all","nv":0,"np":32},{"i":23,"slug":"0xsero","name":"0xSero","topic":"local","why":"Runs frontier open models on his own hardware and writes the recipes down.","yt":null,"ytId":null,"ytSubs":null,"x":"0xSero","av":"https://pbs.twimg.com/profile_images/2111226379521892352/3ICUbYY6_200x200.jpg","mode":"all","nv":0,"np":32},{"i":24,"slug":"alex-ziskind","name":"Alex Ziskind","topic":"local","why":"Benchmarks local models on Macs, PCs and clusters.","yt":"@AZisk","ytId":"UCajiMK_CY9icRhLepS8_3ug","ytSubs":581000,"x":"digitalix","av":"https://pbs.twimg.com/profile_images/1735764125815341056/5KiWd_Ii_200x200.jpg","mode":"ai","nv":817,"np":34},{"i":25,"slug":"digital-spaceport","name":"Digital Spaceport","topic":"local","why":"Home AI servers: GPUs, builds, and what actually runs.","yt":"@DigitalSpaceport","ytId":"UCiaQzXI5528Il6r2NNkrkJA","ytSubs":105000,"x":null,"av":"https://yt3.googleusercontent.com/JjJDFWn5nYfey7NBbjEQCkg-0tYtDFbUPLnKBUePBjWkUuF0cnF6PMtlkRK73FbJzIqLVzf_Prs=s0","mode":"ai","nv":192,"np":0},{"i":26,"slug":"matt-williams","name":"Matt Williams","topic":"local","why":"Ollama founding team. Clear local-model explainers.","yt":"@technovangelist","ytId":"UCHaF9kM2wn8C3CLRwLkC2GQ","ytSubs":96300,"x":"technovangelist","av":"https://pbs.twimg.com/profile_images/1004805401337040899/kAaAiwQW_200x200.jpg","mode":"ai","nv":211,"np":13},{"i":27,"slug":"jeff-geerling","name":"Jeff Geerling","topic":"local","why":"AI on Raspberry Pi, clusters and small hardware. Filtered to AI and computing.","yt":"@JeffGeerling","ytId":"UCR-DXc1voovS8nhAvccRZhg","ytSubs":1100000,"x":"geerlingguy","av":"https://pbs.twimg.com/profile_images/1989562239666282498/5Uk_V_cG_200x200.jpg","mode":"ai","nv":335,"np":55},{"i":28,"slug":"bijan-bowen","name":"Bijan Bowen","topic":"local","why":"Hands-on tests of new open models on local hardware.","yt":"@Bijanbowen","ytId":"UCOCahKBCEUuzDJawM7yN1dg","ytSubs":80900,"x":null,"av":"https://yt3.googleusercontent.com/At-1Kn54AljCodKZ7G9jxzd-1cCR3BknxIqpwGiXckFOXATpOjSyL-ivxtZQPE7RRAtWC8abFA=s0","mode":"ai","nv":430,"np":0},{"i":29,"slug":"sam-witteveen","name":"Sam Witteveen","topic":"local","why":"New open models and agent frameworks, tested in code.","yt":"@samwitteveenai","ytId":"UC55ODQSvARtgSyc8ThfiepQ","ytSubs":140000,"x":"Sam_Witteveen","av":"https://pbs.twimg.com/profile_images/2274169962/jtlg8bxr0w6quu157b7f_200x200.png","mode":"ai","nv":379,"np":21},{"i":30,"slug":"georgi-gerganov","name":"Georgi Gerganov","topic":"local","why":"Creator of llama.cpp and ggml, the engine under most local AI.","yt":null,"ytId":null,"ytSubs":null,"x":"ggerganov","av":"https://pbs.twimg.com/profile_images/1654097134315098113/zCZD0wYz_200x200.jpg","mode":"all","nv":0,"np":46},{"i":31,"slug":"awni-hannun","name":"Awni Hannun","topic":"local","why":"Leads MLX at Apple. Fast local inference on Apple silicon.","yt":null,"ytId":null,"ytSubs":null,"x":"awnihannun","av":"https://pbs.twimg.com/profile_images/829414498893123584/P6JytwO8_200x200.jpg","mode":"all","nv":0,"np":29},{"i":32,"slug":"daniel-han","name":"Daniel Han","topic":"local","why":"Unsloth: fine-tuning and quantizing open models, with bug finds in every release.","yt":null,"ytId":null,"ytSubs":null,"x":"danielhanchen","av":"https://pbs.twimg.com/profile_images/1624054272676532224/UNv4ONME_200x200.jpg","mode":"all","nv":0,"np":31},{"i":33,"slug":"prince-canuma","name":"Prince Canuma","topic":"local","why":"MLX-VLM and day-one ports of new models to the Mac.","yt":null,"ytId":null,"ytSubs":null,"x":"Prince_Canuma","av":null,"mode":"all","nv":0,"np":0},{"i":34,"slug":"ivan-fioravanti","name":"Ivan Fioravanti","topic":"local","why":"Benchmarks every new open model on Apple silicon.","yt":null,"ytId":null,"ytSubs":null,"x":"ivanfioravanti","av":"https://pbs.twimg.com/profile_images/2084888225986826240/-XeooHyq_200x200.jpg","mode":"all","nv":0,"np":26},{"i":35,"slug":"alex-cheema","name":"Alex Cheema","topic":"local","why":"EXO: run big models across the devices you already own.","yt":null,"ytId":null,"ytSubs":null,"x":"alexocheema","av":null,"mode":"all","nv":0,"np":0},{"i":36,"slug":"ahmad-osman","name":"Ahmad Osman","topic":"local","why":"Home GPU clusters and the case for owning your inference.","yt":null,"ytId":null,"ytSubs":null,"x":"TheAhmadOsman","av":"https://pbs.twimg.com/profile_images/1831406651829559296/iMbSOxdP_200x200.jpg","mode":"all","nv":0,"np":15},{"i":37,"slug":"simon-willison","name":"Simon Willison","topic":"agents","why":"The most careful daily notes on LLMs, tools and prompt injection.","yt":null,"ytId":null,"ytSubs":null,"x":"simonw","av":"https://pbs.twimg.com/profile_images/378800000261649705/be9cc55e64014e6d7663c50d7cb9fc75_200x200.jpeg","mode":"all","nv":0,"np":80},{"i":38,"slug":"hamel-husain","name":"Hamel Husain","topic":"agents","why":"Evals: how to tell whether your AI product works.","yt":null,"ytId":null,"ytSubs":null,"x":"HamelHusain","av":"https://pbs.twimg.com/profile_images/1287206199088173057/ixE4fKy1_200x200.jpg","mode":"all","nv":0,"np":27},{"i":39,"slug":"ethan-mollick","name":"Ethan Mollick","topic":"agents","why":"What AI changes for learning and work, tested in practice.","yt":null,"ytId":null,"ytSubs":null,"x":"emollick","av":"https://pbs.twimg.com/profile_images/1601382188712398850/3AAOlqrX_200x200.jpg","mode":"all","nv":0,"np":80},{"i":40,"slug":"boris-cherny","name":"Boris Cherny","topic":"agents","why":"Created Claude Code.","yt":null,"ytId":null,"ytSubs":null,"x":"bcherny","av":"https://pbs.twimg.com/profile_images/1902044548936953856/J2jeik0t_200x200.jpg","mode":"all","nv":0,"np":45},{"i":41,"slug":"dax-raad","name":"Dax Raad","topic":"agents","why":"Builds opencode, an open-source coding agent.","yt":null,"ytId":null,"ytSubs":null,"x":"thdxr","av":"https://pbs.twimg.com/profile_images/1602333093485891584/mmVqjFNI_200x200.jpg","mode":"all","nv":0,"np":41},{"i":42,"slug":"peter-steinberger","name":"Peter Steinberger","topic":"agents","why":"Agentic engineering in the open, shipped daily.","yt":null,"ytId":null,"ytSubs":null,"x":"steipete","av":"https://pbs.twimg.com/profile_images/1131851609774985216/OcsssQ9J_200x200.png","mode":"all","nv":0,"np":15},{"i":43,"slug":"mitchell-hashimoto","name":"Mitchell Hashimoto","topic":"agents","why":"Ghostty and careful notes on working with coding agents.","yt":null,"ytId":null,"ytSubs":null,"x":"mitchellh","av":"https://pbs.twimg.com/profile_images/2082830624415698944/bRo7PB2S_200x200.jpg","mode":"all","nv":0,"np":70},{"i":44,"slug":"omar-khattab","name":"Omar Khattab","topic":"research","why":"DSPy and ColBERT: programming, not prompting, language models.","yt":null,"ytId":null,"ytSubs":null,"x":"lateinteraction","av":"https://pbs.twimg.com/profile_images/1613558765764374528/aZQB6U4b_200x200.jpg","mode":"all","nv":0,"np":6},{"i":45,"slug":"nathan-lambert","name":"Nathan Lambert","topic":"research","why":"Post-training, RLHF and open models, explained.","yt":null,"ytId":null,"ytSubs":null,"x":"natolambert","av":"https://pbs.twimg.com/profile_images/2051823244832083968/Nr0pT8-V_200x200.jpg","mode":"all","nv":0,"np":62},{"i":46,"slug":"lucas-beyer","name":"Lucas Beyer","topic":"research","why":"Vision transformers and sharp research commentary.","yt":null,"ytId":null,"ytSubs":null,"x":"giffmana","av":"https://pbs.twimg.com/profile_images/378800000845687873/37bba4f807fe3a2c644a252f8191338d_200x200.jpeg","mode":"all","nv":0,"np":30},{"i":47,"slug":"jim-fan","name":"Jim Fan","topic":"research","why":"Robotics foundation models and embodied AI.","yt":null,"ytId":null,"ytSubs":null,"x":"DrJimFan","av":"https://pbs.twimg.com/profile_images/1554922493101559808/SYSZhbcd_200x200.jpg","mode":"all","nv":0,"np":64},{"i":48,"slug":"lilian-weng","name":"Lilian Weng","topic":"research","why":"Long, definitive survey posts on agents, RL and safety.","yt":null,"ytId":null,"ytSubs":null,"x":"lilianweng","av":"https://pbs.twimg.com/profile_images/2099378394706448384/pkLvzVZe_200x200.jpg","mode":"all","nv":0,"np":58},{"i":49,"slug":"andrew-ng","name":"Andrew Ng","topic":"foundations","why":"Taught a generation machine learning. DeepLearning.AI.","yt":null,"ytId":null,"ytSubs":null,"x":"AndrewYNg","av":"https://pbs.twimg.com/profile_images/733174243714682880/oyG30NEH_200x200.jpg","mode":"all","nv":0,"np":80},{"i":50,"slug":"francois-chollet","name":"Francois Chollet","topic":"research","why":"Keras and ARC-AGI: what intelligence is and how to measure it.","yt":null,"ytId":null,"ytSubs":null,"x":"fchollet","av":"https://pbs.twimg.com/profile_images/2006806326140350470/Kd5oZv-f_200x200.jpg","mode":"all","nv":0,"np":36},{"i":51,"slug":"chip-huyen","name":"Chip Huyen","topic":"agents","why":"AI engineering and ML systems design.","yt":null,"ytId":null,"ytSubs":null,"x":"chipro","av":"https://pbs.twimg.com/profile_images/1953498653685161984/gMssInYX_200x200.jpg","mode":"all","nv":0,"np":74},{"i":52,"slug":"horace-he","name":"Horace He","topic":"systems","why":"PyTorch compilers and making GPUs go fast.","yt":null,"ytId":null,"ytSubs":null,"x":"cHHillee","av":"https://pbs.twimg.com/profile_images/1747412378751574016/0es3s5Fz_200x200.jpg","mode":"all","nv":0,"np":54},{"i":53,"slug":"tri-dao","name":"Tri Dao","topic":"systems","why":"FlashAttention and Mamba.","yt":null,"ytId":null,"ytSubs":null,"x":"tri_dao","av":"https://pbs.twimg.com/profile_images/1389486324516671496/owa9Z4AC_200x200.jpg","mode":"all","nv":0,"np":32},{"i":54,"slug":"tim-dettmers","name":"Tim Dettmers","topic":"systems","why":"Quantization (QLoRA, bitsandbytes) and which GPU to buy.","yt":null,"ytId":null,"ytSubs":null,"x":"Tim_Dettmers","av":"https://pbs.twimg.com/profile_images/1998431471304876032/BhUl3uX9_200x200.jpg","mode":"all","nv":0,"np":29},{"i":55,"slug":"gpu-mode","name":"GPU MODE","topic":"systems","why":"Lectures on CUDA, Triton and kernels from people who write them.","yt":"@GPUMODE","ytId":"UCJgIbYl6C5no72a0NUAPcTA","ytSubs":34700,"x":"GPU_MODE","av":"https://pbs.twimg.com/profile_images/2065106268281561089/flKZ8X1z_200x200.jpg","mode":"all","nv":124,"np":28},{"i":56,"slug":"asianometry","name":"Asianometry","topic":"systems","why":"How chips are made and why compute is the bottleneck.","yt":"@Asianometry","ytId":"UC1LpsuAUaKoMzzJSEt5WImw","ytSubs":968000,"x":null,"av":"https://yt3.googleusercontent.com/t_WHWwjVFb-837gRQ-ExcsjK1q5Z4dZ-mb4I6ItI4aBo3woGH7MkO7RAKf5kiAZUa3148lgNWw=s0","mode":"ai","nv":354,"np":0},{"i":57,"slug":"steve-brunton","name":"Steve Brunton","topic":"math","why":"Linear algebra, control and physics-informed machine learning.","yt":"@Eigensteve","ytId":"UCm5mt-A4w61lknZ9lCsZtBw","ytSubs":564000,"x":"eigensteve","av":"https://pbs.twimg.com/profile_images/1465359355054723077/IY8OXojh_200x200.jpg","mode":"all","nv":557,"np":9},{"i":58,"slug":"mutual-information","name":"Mutual Information","topic":"math","why":"Probability and RL, explained with care.","yt":"@Mutual_Information","ytId":"UCCcrR0XBH0aWbdffktUBEdw","ytSubs":97200,"x":null,"av":"https://yt3.googleusercontent.com/QKzOz66GtCZU4HyxSqKrBzMy59OXV0Id9ImAZcS74uSCBinsl85I1KVG69Vax_IGxs4-MJJIuw=s0","mode":"all","nv":26,"np":0},{"i":59,"slug":"computerphile","name":"Computerphile","topic":"cs","why":"Computer science from the people at Nottingham who research it.","yt":"@Computerphile","ytId":"UC9-y-6csu5WGm29I7JiwpnA","ytSubs":2640000,"x":null,"av":"https://yt3.googleusercontent.com/ebHMyRfch3u2UTZN1WQJDp9J5U7o38T_WnGkd2QhAIQwBgvozdaOCOnfDMtngtoHWutJvLl4i0c=s0","mode":"all","nv":885,"np":0},{"i":60,"slug":"sebastian-lague","name":"Sebastian Lague","topic":"cs","why":"Coding Adventures: build a thing, explain every idea along the way.","yt":"@SebastianLague","ytId":"UCmtyQOKKmrMVaKuRXz02jbQ","ytSubs":1410000,"x":null,"av":"https://yt3.googleusercontent.com/ytc/AIdro_knyJw3jL_6AGomJmGe3VvfIYxWzZC9Y8z90Liru4G0UHM=s0","mode":"all","nv":280,"np":0},{"i":61,"slug":"ben-eater","name":"Ben Eater","topic":"cs","why":"Build a computer from chips on a breadboard. Then understand all of it.","yt":"@BenEater","ytId":"UCS0N5baNlQWJCUrhCEo8WlA","ytSubs":1400000,"x":null,"av":"https://yt3.googleusercontent.com/ytc/AIdro_kzQ9Fabth3QHyk1YLRHA62goVPgxJdd68G0CPIs0tU3A=s0","mode":"all","nv":132,"np":0},{"i":62,"slug":"core-dumped","name":"Core Dumped","topic":"cs","why":"Operating systems and memory, animated.","yt":"@CoreDumpped","ytId":"UCGKEMK3s-ZPbjVOIuAV8clQ","ytSubs":515000,"x":null,"av":"https://yt3.googleusercontent.com/jfFx60eD6-93bOmWb20Ia6ehdTyiObRpSIzyK1S981LD78uGh8-jbENVajIYOPQCTDMiFfWjzg=s0","mode":"all","nv":49,"np":0},{"i":63,"slug":"reducible","name":"Reducible","topic":"cs","why":"Algorithms and compression, animated in the 3Blue1Brown style.","yt":"@Reducible","ytId":"UCK8XIGR5kRidIw2fWqwyHRA","ytSubs":338000,"x":null,"av":"https://yt3.googleusercontent.com/4Y72xFDw395XHf878UgzsxXc1I-36cgpjgjWaZx7KRP5D3VsMk_bvFoenvYjDRhEGF9Eo2LdPg=s0","mode":"all","nv":23,"np":0},{"i":64,"slug":"spanning-tree","name":"Spanning Tree","topic":"cs","why":"Short, crisp CS explainers.","yt":"@SpanningTree","ytId":"UCDzVUXiTr3hClI-zzCWbYzg","ytSubs":269000,"x":null,"av":"https://yt3.googleusercontent.com/5gIZzhLppqzpN9CbMA0sMax2yimESCzONJnJIrLcB-g-92u8bJzDW7XbRYtPrifa_SusttimIw=s0","mode":"all","nv":29,"np":0},{"i":65,"slug":"fireship","name":"Fireship","topic":"cs","why":"Fast, funny tours of tech. Filtered to AI and CS.","yt":"@Fireship","ytId":"UCsBjURrPoezykLs9EqgamOA","ytSubs":4330000,"x":"fireship_dev","av":"https://pbs.twimg.com/profile_images/1436819851566219267/HEffZjvP_200x200.jpg","mode":"ai","nv":702,"np":14},{"i":66,"slug":"stuff-made-here","name":"Stuff Made Here","topic":"making","why":"Hard engineering projects: CAD, machining, code and a lot of failure.","yt":"@StuffMadeHere","ytId":"UCj1VqrHhDte54oLgPG4xpuQ","ytSubs":4760000,"x":null,"av":"https://yt3.googleusercontent.com/ytc/AIdro_mzfzJ9E1WEzyt1YlPXemHOwGBBnVCKzXva_mzmS5qqMg=s0","mode":"all","nv":41,"np":0},{"i":67,"slug":"cnc-kitchen","name":"CNC Kitchen","topic":"making","why":"Stefan tests 3D printing claims with real measurements.","yt":"@CNCKitchen","ytId":"UCiczXOhGpvoQGhOL16EZiTg","ytSubs":767000,"x":null,"av":"https://yt3.googleusercontent.com/356qAi2EL_vfUQsQNFlYWP_x2TK_Ng3MOfhhO_nPAHd91JguvNA61dZUf1LTxTuDOclgxaVD-w=s0","mode":"all","nv":225,"np":0},{"i":68,"slug":"teaching-tech","name":"Teaching Tech","topic":"making","why":"Printer setup, calibration and CAD for makers.","yt":"@TeachingTech","ytId":"UCbgBDBrwsikmtoLqtpc59Bw","ytSubs":579000,"x":null,"av":"https://yt3.googleusercontent.com/ytc/AIdro_kRtIOW7AUMB72XCkUCJwdrWnF3C8bNoXIHOJFqizwI60g=s0","mode":"all","nv":553,"np":0},{"i":69,"slug":"thomas-sanladerer","name":"Thomas Sanladerer","topic":"making","why":"3D printing materials, machines and honest reviews.","yt":"@MadeWithLayers","ytId":"UCb8Rde3uRL1ohROUVg46h1A","ytSubs":535000,"x":null,"av":"https://yt3.googleusercontent.com/b7R_4be18afRCpEPUe4R_XVGNW2qWfBGVrga1e-BmQPjmqEL0mg0L1er1sVSXwXSIgKwS1g0l_Y=s0","mode":"all","nv":477,"np":0},{"i":70,"slug":"maker-s-muse","name":"Maker's Muse","topic":"making","why":"Design for 3D printing: tolerances, mechanisms, print-in-place.","yt":"@MakersMuse","ytId":"UCxQbYGpbdrh-b2ND-AfIybg","ytSubs":1180000,"x":null,"av":"https://yt3.googleusercontent.com/ytc/AIdro_kmQ8LEMA74i3vrUn2JajOz5NQ2yTK9zREX9eNOpGtf8nY=s0","mode":"all","nv":585,"np":0},{"i":71,"slug":"zack-freedman","name":"Zack Freedman","topic":"making","why":"Wearable and printed gadgets, designed in the open.","yt":"@ZackFreedman","ytId":"UCUW49KGPezggFi0PGyDvcvg","ytSubs":681000,"x":"ZackFreedman","av":"https://pbs.twimg.com/profile_images/1329252128775417862/QkOG9h-K_200x200.jpg","mode":"all","nv":96,"np":5},{"i":72,"slug":"product-design-online","name":"Product Design Online","topic":"making","why":"Learn Fusion CAD from zero, one project at a time.","yt":"@ProductDesignOnline","ytId":"UCooViVfi0DaWk_eqxIXXiOQ","ytSubs":391000,"x":null,"av":"https://yt3.googleusercontent.com/IlO8XRNxytRs1jhga1aX7UqXdl_PTgkYE8GM_1XIUJTY7j-mpWO5bOe80gIgUmYvKNzJXRGBgJs=s0","mode":"all","nv":242,"np":0},{"i":73,"slug":"mangojelly-solutions","name":"MangoJelly Solutions","topic":"making","why":"The best FreeCAD tutorials. Free software, free lessons.","yt":"@MangoJellySolutions","ytId":"UCUWhaOxsRk_5oPPq00_Y7Dw","ytSubs":128000,"x":null,"av":"https://yt3.googleusercontent.com/Kg6Dh_MDlSwO2cblSDgq0gCu2gRURXKiU9V-ktj_q8vco6uVn99y9hi1A6kzgJscy8sHCxdvDg=s0","mode":"all","nv":613,"np":0},{"i":74,"slug":"zoo","name":"Zoo","topic":"making","why":"Text-to-CAD: AI that writes real, editable CAD models.","yt":null,"ytId":null,"ytSubs":null,"x":"zoodotdev","av":"https://pbs.twimg.com/profile_images/1938298729066577920/a0gq-Y4h_200x200.jpg","mode":"all","nv":0,"np":59},{"i":75,"slug":"anthropic","name":"Anthropic","topic":"labs","why":"Claude, interpretability and safety research.","yt":"@anthropic-ai","ytId":"UCrDwWp7EBBv4NwvScIpBDOA","ytSubs":834000,"x":"AnthropicAI","av":"https://pbs.twimg.com/profile_images/1798110641414443008/XP8gyBaY_200x200.jpg","mode":"all","nv":119,"np":80},{"i":76,"slug":"claude","name":"Claude","topic":"labs","why":"Claude product releases and how-tos.","yt":null,"ytId":null,"ytSubs":null,"x":"claudeai","av":"https://pbs.twimg.com/profile_images/1950950107937185792/QOfEjFoJ_200x200.jpg","mode":"all","nv":0,"np":39},{"i":77,"slug":"openai","name":"OpenAI","topic":"labs","why":"Model launches and research.","yt":"@OpenAI","ytId":"UCXZCJLdBC09xxGZ6gcdrc6A","ytSubs":2110000,"x":"OpenAI","av":"https://pbs.twimg.com/profile_images/1885410181409820672/ztsaR0JW_200x200.jpg","mode":"all","nv":537,"np":80},{"i":78,"slug":"google-deepmind","name":"Google DeepMind","topic":"labs","why":"AlphaFold, Gemini and the science behind them.","yt":"@googledeepmind","ytId":"UCP7jMXSY2xbc3KCAE0MHQ-A","ytSubs":941000,"x":"GoogleDeepMind","av":"https://pbs.twimg.com/profile_images/1695024885070737408/-M-HSH5P_200x200.jpg","mode":"all","nv":246,"np":78},{"i":79,"slug":"hugging-face","name":"Hugging Face","topic":"labs","why":"The home of open models, with courses to match.","yt":"@HuggingFace","ytId":"UCHlNU7kIZhRgSbhHvFoy72w","ytSubs":150000,"x":"huggingface","av":"https://pbs.twimg.com/profile_images/1991559933473497089/mbrRS49P_200x200.jpg","mode":"all","nv":300,"np":9},{"i":80,"slug":"qwen","name":"Qwen","topic":"labs","why":"Open-weight models that run well locally.","yt":null,"ytId":null,"ytSubs":null,"x":"Alibaba_Qwen","av":"https://pbs.twimg.com/profile_images/2064231947149377536/Ab70PxT5_200x200.jpg","mode":"all","nv":0,"np":53},{"i":81,"slug":"deepseek","name":"DeepSeek","topic":"labs","why":"Open frontier models and the papers behind them.","yt":null,"ytId":null,"ytSubs":null,"x":"deepseek_ai","av":"https://pbs.twimg.com/profile_images/1717417613775757312/Uk1zNOj4_200x200.jpg","mode":"all","nv":0,"np":71},{"i":82,"slug":"ollama","name":"Ollama","topic":"local","why":"One command to run a model on your own machine.","yt":null,"ytId":null,"ytSubs":null,"x":"ollama","av":"https://pbs.twimg.com/profile_images/2045273960666071040/rCjCIwiC_200x200.jpg","mode":"all","nv":0,"np":42},{"i":83,"slug":"lm-studio","name":"LM Studio","topic":"local","why":"A desktop app for local models, GGUF and MLX.","yt":null,"ytId":null,"ytSubs":null,"x":"lmstudio","av":"https://pbs.twimg.com/profile_images/1755060270173429760/4WVc54_p_200x200.jpg","mode":"all","nv":0,"np":52},{"i":84,"slug":"unsloth","name":"Unsloth","topic":"local","why":"Fine-tune open models faster on less memory.","yt":null,"ytId":null,"ytSubs":null,"x":"UnslothAI","av":"https://pbs.twimg.com/profile_images/2036313340616777728/OLALWB2__200x200.jpg","mode":"all","nv":0,"np":80},{"i":85,"slug":"vllm","name":"vLLM","topic":"systems","why":"The open serving engine most providers run.","yt":null,"ytId":null,"ytSubs":null,"x":"vllm_project","av":"https://pbs.twimg.com/profile_images/1774187681746182144/N_5NJ8B1_200x200.jpg","mode":"all","nv":0,"np":80}],"meta":{"built":1794089672,"videos":16205,"posts":2660,"people":86,"shorts":687,"byTopic":{"foundations":1612,"local":629,"agents":2347,"research":1781,"math":906,"cs":3453,"systems":959,"making":3399,"labs":1119},"postsByTopic":{"foundations":264,"local":348,"agents":544,"research":532,"math":116,"cs":151,"systems":240,"making":135,"labs":330}},"videos":[["1QNsdr-Qx_I","Introducing GPT-6 Astra: the most intelligent and aligned model in the world.",163,3700000,1788825600,77,8,3,92],["HK6y8DAPN_0","Introducing Sora — OpenAI’s text-to-video model",624,4200000,1731024000,77,8,3,88],["bOC3DisEOfg","Introducing GPT-6 Astra for developers",204,1200000,1788825600,77,8,3,86],["yyvVUEPSCu0","Introducing workspace agents in ChatGPT",71,3400000,1778198400,77,8,2,84],["xIEzjwReF2s","Germany: This is the new ChatGPT Voice, powered by GPT Live",194,8000000,1792886400,77,8,1,79],["44eFf-tRiSg","Intro to Agent Builder",330,1700000,1762560000,77,2,2,79],["boJG84Jcf-4","Introducing GPT-5",90,621000,1762560000,77,8,3,79],["ZXEAuZOSLs0","What Codex Unlocks for Ramp",62,14000000,1775606400,77,2,1,78],["4_P7R8Fjrg8","What Codex Unlocks for Ryan Hendler, dev at me&u",72,15000000,1775606400,77,2,1,78],["kopoLzvh5jY","Multi-Agent Hide and Seek",178,11000000,1573171200,77,3,2,78],["_NLHFoVNlbg","Stanford CS230 | Autumn 2025 | Lecture 1: Introduction to Deep Learning",3617,760000,1762560000,7,0,3,91],["SQ3fZ1sAqXI","Stanford CS336 Language Modeling from Scratch | Spring 2025 | Lecture 1: Overview and Tokenization",4739,800000,1762560000,7,0,3,91],["rmVRLeJRkl4","Stanford CS224N: NLP with Deep Learning | Winter 2021 | Lecture 1 - Intro & Word Vectors",5067,818000,1636329600,7,0,3,85],["kJLiOGle3Lw","Stanford Webinar - Agentic AI: A Progression of Language Model Usage",3426,699000,1762560000,7,2,2,84],["kV1ru-Inzl4","Stanford EE364A Convex Optimization I Stephen Boyd I 2023 I Lecture 1",4707,216000,1731024000,7,4,3,84],["WsvFL-LjA6U","Stanford CS234 Reinforcement Learning I Introduction to Reinforcement Learning I 2024 I Lecture 1",4748,204000,1731024000,7,0,3,83],["mE7IDf2SmJg","Stanford CS25: V3 I Retrieval Augmented Language Models",4767,205000,1731024000,7,3,3,83],["V1tINV2-9p4","Stanford CS149 I Parallel Computing I 2023 I Lecture 1 - Why Parallelism? Why Efficiency?",4342,153000,1731024000,7,6,3,82],["vkUNH9r6UCI","Stanford Lecture - Don Knuth:  The Analysis of Algorithms (2015, recreating 1969)",3265,156000,1510099200,7,5,3,79],["qglA5eql38c","Course Overview - Technical Fundamentals of Generative AI",88,3100000,1783468800,7,0,1,78],["CEvIs9y1uog","Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic",982,1500000,1767830400,18,2,3,91],["D7_ipDqhtwk","How We Build Effective Agents: Barry Zhang, Anthropic",909,584000,1762560000,18,2,3,90],["eZ8WWZzoaR0","Homa: The End of TCP for AI Clusters — John Ousterhout, Stanford",1128,309000,1791417600,18,6,3,88],["C_GG5g38vLU","Harnesses in AI: A Deep Dive — Tejas Kumar, IBM",1227,336000,1780876800,18,2,3,88],["kQmXtrmQ5Zg","Building Agents with Model Context Protocol - Full Workshop with Mahesh Murag of Anthropic",6252,360000,1762560000,18,2,3,87],["am_oeAoUhew","Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI",2781,260000,1778198400,18,2,3,86],["8kMaTybvDUw","12-Factor Agents: Patterns of reliable LLM applications — Dex Horthy, HumanLayer",1026,297000,1762560000,18,2,3,86],["Ib5GBkD555M","Why Software Factories Fail",1158,636000,1786147200,18,2,2,85],["TRjq7t2Ms5I","Building Production-Ready RAG Applications: Jerry Liu",1115,412000,1731024000,18,2,3,85],["Sir59K8ZDPU","Why Agentic Systems Need Ontologies — Frank Coyle, UC Berkeley",1278,500000,1786147200,18,2,2,84],["K75j8MkwgJ0","Optimize Your AI - Quantization Explained",730,555000,1762560000,26,1,3,91],["3UQ7GY9hNwk","Fine Tune a model with MLX for Ollama",520,179000,1731024000,26,0,2,75],["aGwb1KLmtog","7. Embeddings in Depth - Part of the Ollama Course",627,38000,1731024000,26,0,3,72],["GMlSFIp1na0","Upgrade Your AI Using Web Search - The Ollama Course",492,100000,1731024000,26,2,2,71],["8r9Kit3lKXE","5. Comparing Quantizations of the Same Model - Ollama Course",629,34000,1731024000,26,1,3,71],["dMY3dBLojTk","Fast Fine Tuning with Unsloth",989,56000,1762560000,26,0,2,70],["VDuA5xbkEjo","Using Ollama and N8N for AI Automation",823,80000,1731024000,26,2,2,70],["jLVl5V8roMU","Have You Picked the Wrong AI Agent Framework?",790,90000,1731024000,26,2,2,70],["BCfCdTp-fdM","Is MLX the best Fine Tuning Framework?",1148,47000,1762560000,26,1,2,69],["IdPdwQdM9lA","How does function calling with tools really work?",609,23000,1731024000,26,2,3,69],["IZAlq-V19U8","Claude Code’s Bitter Lesson: Prompts, Harnesses, Mods, and Agents — Thariq Shihipar, Anthropic",5672,149000,1791417600,17,2,3,87],["3uSI8q_RN-o","The Inference Frontier: from 100 to 10,000 tokens per second — Sean Lie, Cerebras CTO",2643,413000,1788825600,17,6,2,86],["cFx9Z3ZXca0","Why I couldn't build Jev at OpenAI — Diogo Almeida, TypeSafe Co-founder & CEO",8542,357000,1791417600,17,2,2,85],["YHiqVRxGViM","AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic Labs",5121,160000,1791676800,17,3,2,80],["fpAthTtha8c","Cooking with OpenAI’s Research Chief: AGI, o1, Evals, and Scaling Laws — Mark Chen",2478,153000,1783468800,17,3,2,80],["h5dlIPM0X18","Why AI Labs With Unlimited GPUs Still Fail — Anjney Midha, AMP",3637,175000,1783468800,17,3,2,80],["79mIutht1f4","🔬 Why Transformers Hit a Wall the Moment Physics Shows Up — Anima Anandkumar, Caltech",5012,121000,1788825600,17,3,2,79],["KpOW9Pk4BUs","Simulating Humanity: from Generative Agents to 8 Billion Digital Twins — Joon Sung Park, Simile AI",4261,128000,1788825600,17,2,2,79],["kog7mwsDqnk","Recursive Language Models — Alex Zhang, MIT PhD",6205,99000,1791417600,17,3,2,78],["z9OkBD2-MDU","OpenAI’s New Agent Stack: Computer Use, Decisions API, UltraFast, Dots—Ari Weinstein & Nikunj Handa",2424,105000,1791417600,17,8,2,78],["FL6mI_Br-mc","Create your own voices with Gemini 3.8 text-to-speech",60,6000000,1791417600,78,8,2,86],["U0aToL5C-bQ","AlphaGenome Atlas: Understanding the human genome",217,4900000,1791417600,78,3,2,86],["LVEfFyijQ2c","Applying Gemini 4 for large-scale codebase maintenance in Chromium",155,3300000,1792281600,78,2,2,85],["aSY_vFFmkW0","Generating novel scientific hypotheses with Co-Scientist",385,6200000,1780876800,78,3,2,85],["sFfPkUJd_Jk","Genie 4: A new generation of world simulation",148,614000,1793145600,78,8,3,82],["2pWv7GOvuf0","RL Course by David Silver - Lecture 1: Introduction to Reinforcement Learning",5293,1900000,1446940800,78,0,3,81],["tBjgCj_dGZM","The mathematics of AI uncertainty",2682,400000,1788825600,78,4,3,79],["ftkic7Tdpcg","Using Gemini 4 Argon to protect against Cyber attackers",128,1000000,1792281600,78,5,2,78],["9PICEmtfpWU","Teaching a humanoid robot to play cornhole with Gemini 4 Argon",104,2900000,1792281600,78,7,1,78],["zzXyPGEtseI","Is human data enough? | David Silver",2971,459000,1762560000,78,3,3,78],["ZPsLhvgU8kc","FreeCAD 1.0 Ultimate Beginners Crash Course in 1 hour for 3D printing   CAD Lesson / Tutorial",3759,585000,1762560000,73,7,3,85],["9cqs3oTzpac","FreeCAD 1.1 Ultimate Beginners Crash Course in 1 hour for 3D printing CAD Lesson / Tutorial",3914,273000,1780876800,73,7,3,82],["NXN7TOg3kj4","FreeCAD 0.20 For Beginners | 1 | Setup , preferences and navigation | Easy Tutorial for Learning CAD",1444,636000,1699401600,73,7,3,80],["uf32NAq-Rsk","FreeCAD 0.20 For Beginners | 2 | Create a simple model and export to STL in Part Design",2391,453000,1667865600,73,7,3,78],["p2yDGlv5wI0","FreeCAD 1.0 Assembly in 30 minutes Beginners Crash Course / Tutorial 2025",2024,150000,1762560000,73,7,3,77],["WdTyFHY-5r4","FreeCAD 26.3 RC1 IS HERE! AND IT'S BIG!  |  I Take a Deep Look Into the Release Candidate",578,324000,1791676800,73,7,2,76],["qVp4GX_z110","FreeCAD 1.1 Assembly in 30 minutes Beginners Crash Course / Tutorial 2026",2010,72000,1780876800,73,7,3,73],["fJQvFsGXz-c","FreeCAD 0.20 For Beginners | 4 | Tracing Photo / Image to Build a Simple Model",1966,210000,1667865600,73,7,3,73],["DFTGt_-jeos","FreeCAD 0.20 For Beginners | 3 | Refining Part Design Model and Output To Technical Drawing",2808,215000,1667865600,73,7,3,73],["M-xUFwzAt6A","What's New in FreeCAD 26.3?   It's not v1.2 anymore!",704,134000,1786147200,73,7,2,71],["_yrw6c5gw3E","M5 Ultra vs 2 DGX Sparks… The Number You're Not Looking At",1238,1100000,1791417600,24,6,2,81],["ujs0_cpAnaw","I Gave Local AI and the Cloud the Exact Same Job",851,1000000,1788825600,24,1,2,80],["FE1Uyhg5hjw","AMD Says 2 Ryzen AI Halos Can Run a 400B Model... I Tested It",1062,842000,1786147200,24,1,2,79],["RkzCAaIV_cQ","This Shouldn’t Be Able to Run 120B Locally",733,958000,1775606400,24,1,2,79],["GBR6pHZ68Ho","M4 Mac Mini CLUSTER 🤯",1086,849000,1762560000,24,6,2,77],["gnincqWDt5w","Framework 13 Pro Shows Why Apple Solders Your Memory",1296,527000,1786147200,24,6,2,76],["c6u87wtQzTw","192GB of VRAM in One PC… The Cheap Way",935,497000,1786147200,24,1,2,76],["-aEHitayNts","I Tested the Cheapest Path to 96GB of VRAM",1188,599000,1775606400,24,1,2,76],["Cyn_Dm05_eU","I Ran Claude Code for FREE… Here's How",905,402000,1772928000,24,2,2,73],["Ze5XLooTt6g","I Thought DGX Spark Was Slower… Until I Changed ONE Thing",904,396000,1770508800,24,6,2,73],["I0-izyq6q5s","How Do Computers Remember?",1172,6200000,1636329600,60,5,3,81],["U4ogK0MIzqk","Coding Adventure: Chess",1762,4200000,1636329600,60,5,3,79],["QZwneRb-zqA","Exploring How Computers Work",1092,4000000,1636329600,60,5,3,79],["rSKMYc1CQHE","Coding Adventure: Simulating Fluids",2872,2600000,1699401600,60,5,3,77],["qyNWDzO6kKQ","I Made a Pen Plotter to Draw Maps of the Moon",1855,663000,1791417600,60,7,3,75],["kOkfC5fLfgE","Coding Adventure: Rendering Fluids",3521,954000,1762560000,60,5,3,75],["hfMk-kjRv4c","I Tried Coding a Neural Network to Identify Doodles",3291,2200000,1667865600,60,0,3,75],["X-iSQQgOd1A","Coding Adventure: Ant and Slime Simulations",1074,2100000,1636329600,60,5,3,75],["-L-WgKMFuhE","A* Pathfinding (E01: algorithm explanation)",699,2200000,1446940800,60,5,3,75],["SO83KQuuZvg","Coding Adventure: Rendering Text",4254,1200000,1731024000,60,5,3,74],["quh7z1q7-uc","Building LLMs from the Ground Up: A 3-hour Coding Workshop",9910,175000,1731024000,5,0,3,80],["kPGTx4wcm_w","Developing an LLM: Building, Training, Finetuning",3526,153000,1731024000,5,0,3,80],["-Ll8DtpNtvk","Build an LLM from Scratch 3: Coding attention mechanisms",8141,79000,1762560000,5,0,3,78],["yAcWnfsZhzo","Build an LLM from Scratch 1: Set up your code environment",1262,245000,1762560000,5,0,2,78],["KT4n-z_4QJU","OpenAI Astra and Recurrent Depth / Looped Transformers",1684,125000,1788825600,5,3,2,76],["rNlULI-zGcw","The Big LLM Architecture Comparison",5197,48000,1762560000,5,0,3,75],["4yNswvhPWCQ","Build an LLM from Scratch 7: Instruction Finetuning",6364,50000,1762560000,5,0,3,75],["Zar2TJv-sE0","Build an LLM from Scratch 5: Pretraining on Unlabeled Data",9404,41000,1762560000,5,0,3,74],["YSAkgEarBGE","Build an LLM from Scratch 4: Implementing a GPT model from Scratch To Generate Text",6337,43000,1762560000,5,0,3,74],["341Rb8fJxY0","Build an LLM from Scratch 2: Working with text data",5281,124000,1762560000,5,0,2,74],["IpGxLWOIZy4","A Friendly Introduction to Machine Learning",1849,977000,1478563200,9,0,3,80],["2-Ol7ZB0MmU","A friendly introduction to Convolutional Neural Networks and Image Recognition",1928,670000,1510099200,9,0,3,78],["BR9h47Jtqyw","A friendly introduction to Deep Learning and Neural Networks",2000,725000,1510099200,9,0,3,78],["UNmqTiOnRfg","A friendly introduction to Recurrent Neural Networks",1364,618000,1510099200,9,0,3,77],["kqSzLo9fenk","A friendly introduction to Bayes Theorem and Hidden Markov Models",1966,526000,1541635200,9,0,3,76],["ZspR5PZemcs","How does Netflix recommend movies? Matrix Factorization",1966,401000,1541635200,9,0,3,75],["9r7FIXEAGvs","Shannon Entropy and Information Gain",1276,227000,1510099200,9,4,3,71],["g-Hb26agBFg","Principal Component Analysis (PCA)",1594,469000,1573171200,9,0,2,69],["UPtG_38Oq8o","The math behind Attention: Keys, Queries, and Values matrices",2176,397000,1699401600,9,0,2,68],["aDW44NPhNw0","Machine Learning: Testing and Error Metrics",2683,119000,1510099200,9,0,3,67],["iDulhoQ2pro","Attention Is All You Need",1627,815000,1541635200,11,0,3,80],["bAWV_yrqx4w","[GRPO Explained] DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models",4140,176000,1762560000,11,0,3,76],["TrdevFK_am4","An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Paper Explained)",1796,400000,1604793600,11,0,3,75],["9dSkvxS2EB0","Mamba: Linear-Time Sequence Modeling with Selective State Spaces (Paper Explained)",2440,179000,1731024000,11,3,3,73],["SY5PvZrJhLE","GPT-3: Language Models are Few-Shot Learners (Paper Explained)",3870,219000,1604793600,11,3,3,72],["7NNxK3CqaDk","Flow Matching for Generative Modeling (Paper Explained)",3376,113000,1731024000,11,0,3,71],["nOBm4aYEYR4","What is Q-Learning (back to basics)",2744,118000,1731024000,11,0,3,71],["3N3Bl5AA5QU","This is a game changer! (AlphaTensor by DeepMind explained)",3307,185000,1667865600,11,4,3,71],["W-O7AZNzbzQ","DDPM - Diffusion Models Beat GANs on Image Synthesis (Machine Learning Research Paper Explained)",3274,169000,1636329600,11,0,3,70],["T9XSU0pKX2E","OpenAI CLIP: ConnectingText and Images (Paper Explained)",2887,182000,1636329600,11,0,3,70],["URtF_UHYBSo","The Elegant Math Behind Machine Learning",6792,563000,1731024000,13,4,3,80],["bIZB1hIJ4u8","The Geometric Deep Learning Blueprint — Michael Bronstein",12803,422000,1636329600,13,0,3,75],["sJXn4Cl4oww","This is why Deep Learning is really weird.",7598,653000,1731024000,13,0,2,74],["DtePicx_kFY","He Co-Invented the Transformer. Now: Continuous Thought Machines [Llion Jones / Luke Darlow]",4360,114000,1765152000,13,3,3,73],["YpFaPKOeNME","NEURAL NETWORKS ARE WEIRD! - Neel Nanda (DeepMind)",13357,112000,1762560000,13,3,3,73],["M-jTeBCEGHc","The Real Reason Huge AI Models Actually Work [Prof. Andrew Wilson]",7429,82000,1762560000,13,3,3,71],["kuvFoXzTK3E","Prof. Chris Bishop's NEW Deep Learning Textbook!",4980,135000,1731024000,13,0,3,71],["AREWYbVtX64","Intelligence is collective, not artificial — Prof. Michael I. Jordan (UC Berkeley / Inria)",4630,54000,1780876800,13,3,3,70],["DP454c1K_vQ","The Man Who Invented Modern AI (Before Everyone Else) — Jürgen Schmidhuber",5979,101000,1731024000,13,0,3,70],["dkpDjd2nHgo","Mystery of Entropy FINALLY Solved After 50 Years? (STEPHEN WOLFRAM)",5047,502000,1699401600,13,4,2,70],["7xTGNNLPyMI","Deep Dive into LLMs like ChatGPT",12684,10000000,1762560000,0,0,3,79],["kCc8FmEb1nY","Let's build GPT: from scratch, in code, spelled out.",6980,8200000,1699401600,0,0,3,73],["zjkBMFhNj_g","[1hr Talk] Intro to Large Language Models",3588,4300000,1731024000,0,0,3,72],["VMj-3S1tku0","The spelled-out intro to neural networks and backpropagation: building micrograd",8752,4400000,1667865600,0,0,3,69],["EWvNQjAaOHw","How I use LLMs",7872,2800000,1762560000,0,2,2,65],["l8pRSuU81PU","Let's reproduce GPT-2 (124M)",14486,1200000,1731024000,0,0,3,64],["zduSFxRajkE","Let's build the GPT Tokenizer",8015,1200000,1731024000,0,0,3,64],["PaCmpygFfXo","The spelled-out intro to language modeling: building makemore",7065,1300000,1667865600,0,0,3,62],["P6sfmUTpUmc","Building makemore Part 3: Activations & Gradients, BatchNorm",6958,569000,1667865600,0,0,3,57],["TCH_1BHY58I","Building makemore Part 2: MLP",4540,613000,1667865600,0,0,3,57],["alfdI7S6wCY","MIT Introduction to Deep Learning (2025) | 6.S191",4166,1000000,1762560000,8,0,3,79],["QDX-1M5Nj7s","MIT Introduction to Deep Learning (2023) | 6.S191",3492,2000000,1699401600,8,0,3,78],["ErnWZxJovaM","MIT Introduction to Deep Learning (2024) | 6.S191",4198,1000000,1731024000,8,0,3,77],["II4giR4vOOo","MIT Introduction to Deep Learning | 6.S191",3376,423000,1775606400,8,0,3,75],["njKP3FqW3Sk","MIT 6.S191 (2020): Introduction to Deep Learning",3172,1200000,1604793600,8,0,3,75],["GvezxUdLrEk","MIT 6.S191 (2025): Recurrent Neural Networks, Transformers, and Attention",3694,305000,1762560000,8,0,3,72],["ySEx_Bqxvvo","MIT 6.S191 (2023): Recurrent Neural Networks, Transformers, and Attention",3770,699000,1699401600,8,0,3,72],["7sB052Pz0sQ","MIT Introduction to Deep Learning (2022) | 6.S191",2941,632000,1667865600,8,0,3,71],["oGpzWAlP5p0","MIT 6.S191 (2025): Convolutional Neural Networks",3664,231000,1762560000,8,0,3,70],["dqoEU9Ac3ek","MIT 6.S191 (2024): Recurrent Neural Networks, Transformers, and Attention",3691,307000,1731024000,8,0,3,70],["aR20FWCCjAs","Ilya Sutskever – We're moving from the age of scaling to the age of research",5764,1600000,1765152000,14,3,3,79],["lXUZvyajciY","Andrej Karpathy — “We’re summoning ghosts, not building animals”",8768,1400000,1762560000,14,3,3,78],["X50zezLFWWI","Ajeya Cotra – \"This might be the clearest warning shot we ever get\"",8433,1700000,1788825600,14,3,2,75],["xmkSf5IS-zw","How GPT, Claude, and Gemini are actually trained and served – Reiner Pope",8021,514000,1778198400,14,6,3,74],["Q8Fkpi18QXU","Terence Tao – How the world’s top mathematician uses AI",5024,514000,1775606400,14,4,3,73],["u15N3l4RT80","The OpenAI/Hugging Face attack, clearly explained",1481,1100000,1788825600,14,5,2,72],["oIk3R-sMX5o","Chip design from the bottom up – Reiner Pope",4820,341000,1780876800,14,6,3,71],["Hrbq66XqtCo","Jensen Huang – Will Nvidia’s moat persist?",6193,998000,1778198400,14,6,2,71],["mDG_Hx3BSUE","Dylan Patel — The single biggest bottleneck to scaling AI compute",9045,332000,1775606400,14,6,3,71],["n1E9IZfvGMA","Dario Amodei — “We are near the end of the exponential”",8540,1200000,1772928000,14,3,2,71],["jkrNMKz9pWU","A Hackers' Guide to Language Models",5473,580000,1699401600,6,0,3,78],["8SF_h3xF3cE","Practical Deep Learning for Coders: Lesson 1",4976,612000,1667865600,6,0,3,78],["XfoYk_Z5AkI","Lesson 1: Deep Learning 2019 - Image classification",6012,423000,1573171200,6,0,3,76],["IPBSB1HLNLo","Lesson 1: Deep Learning 2018",5195,413000,1541635200,6,0,3,76],["CzdWqFTmn0Y","Intro to Machine Learning: Lesson 1",4661,341000,1541635200,6,0,3,75],["_QUEXsHfsA0","Lesson 1 - Deep Learning for Coders (2020)",4951,336000,1604793600,6,0,3,74],["F4tvM4Vb3A0","Lesson 2: Practical Deep Learning for Coders 2022",4602,247000,1667865600,6,0,3,73],["_7rMfsA24Ls","Lesson 9: Deep Learning Foundations to Stable Diffusion",8116,182000,1667865600,6,0,3,71],["ccMHJeQU4Qw","Lesson 2: Deep Learning 2019 - Data cleaning and production; SGD from scratch",7128,187000,1573171200,6,0,3,71],["Th_ckFbc6bI","Lesson 1: Practical Deep Learning for Coders",5926,206000,1510099200,6,0,3,71],["Lu56xVlZ40M","OpenAI Plays Hide and Seek…and Breaks The Game! 🤖",362,11000000,1573171200,12,3,2,78],["SsJ_AusntiU","This AI Learned Boxing…With Serious Knockout Power! 🥊",363,4400000,1636329600,12,3,2,74],["7NF3CdXkm68","NVIDIA Just Solved The Hardest Problem in Physics Simulation!",469,1600000,1762560000,12,5,2,73],["54YvCE8_7lM","Is Simulating Soft and Bouncy Jelly Possible? 🦑",350,2900000,1604793600,12,5,2,71],["7SM816P5G9s","Is a Realistic Honey Simulation Possible? 🍯",276,2200000,1573171200,12,5,2,69],["_WjU5d26Cc4","Intel Just Changed Computer Graphics Forever!",399,611000,1762560000,12,6,2,67],["2Bw5f4vYL98","How Well Can DeepMind's AI Learn Physics? ⚛",438,1600000,1604793600,12,3,2,67],["u5wtoH0_KuA","This AI Does Nothing In Games…And Still Wins!",417,1400000,1604793600,12,3,2,67],["tfb6aEUMC04","OpenAI Five Beats World Champion DOTA2 Team 2-0! 🤖",691,1500000,1573171200,12,8,2,67],["V1eYniJ0Rnk","Google DeepMind's Deep Q-learning playing Atari Breakout!",103,1400000,1446940800,12,3,2,67],["LuhJEEJQgUM","Lecture 1 How to profile CUDA kernels in PyTorch",3373,61000,1731024000,55,6,3,78],["KK6TWf4Pp6A","Lecture 117:Building a GPU Compiler from Scratch",3241,14000,1791676800,55,5,3,74],["DdTsX6DQk24","Lecture 14: Practitioners Guide to Triton",4903,24000,1731024000,55,6,3,72],["JwUcZwPOCpA","Lecture 36: CUTLASS and Flash Attention 3",6556,12000,1762560000,55,6,3,71],["4sgKnKbR-WE","Lecture 3: Getting Started With CUDA for Python Programmers",4676,20000,1731024000,55,6,3,71],["NQ-0D5Ti2dc","Lecture 2 Ch1-3 PMPP book",3146,19000,1731024000,55,6,3,71],["1E8GDR8QXKw","Lecture 48: The Ultra Scale Playbook",11028,11000,1762560000,55,6,3,70],["hQ9GPnV0-50","Lecture 23: Tensor Cores",6470,16000,1731024000,55,6,3,70],["lTmYrKwjSOU","Lecture 4 Compute and Memory Basics",3415,16000,1731024000,55,6,3,70],["VPslgC9piIw","Lecture 80: How FlashAttention 4 Works",4509,10000,1762560000,55,6,3,69],["T-Z3GmM20JM","Complete beginner's guide to 3D printing - Assembly, tour, slicing, levelling and first prints",1467,2900000,1604793600,68,7,3,78],["gmePlcU0TRw","Achieve true 3D printing with non planar slicing",848,1400000,1573171200,68,7,3,73],["UU6tWhV010M","Resin 3D printing beginners step by step guide",1429,1200000,1667865600,68,7,3,72],["I3xMHjiYENk","Linear rails guide and test on Ender 3 and Ender 5",1025,748000,1604793600,68,7,3,69],["_ramiM3KHYE","CoreXY explained: Comparison + strengths & weaknesses",752,660000,1604793600,68,7,3,69],["3yIebnVjADM","How to tune your slicer settings featuring Ender 3 - UPDATE IN DESCRIPTION",672,742000,1541635200,68,7,3,69],["n3yK0lJ8TWM","Linear advance guide - Free and easy print quality improvement",728,603000,1573171200,68,7,3,68],["vl0FT339jfc","Eliminate seams in your 3D prints with scarf joint seams - An exciting development!",709,964000,1731024000,68,7,2,67],["cquTCpz1V74","Orca Slicer getting started guide: A slicer for all of your 3D printers",1057,458000,1699401600,68,7,3,67],["NVtsrisYtUE","Over centre mechanisms will make your designs more effective - Guide with examples",780,450000,1699401600,68,7,3,67],["fRqwYsfiME8","Incredible Rolling Objects which aren't Spheres!",446,13000000,1541635200,70,4,2,78],["Le-PVXUCl0M","The Geneva Mechanism",436,6000000,1510099200,70,7,2,76],["r9ZTWQtzk2Y","This Robot Drives on Wobbly Wheels (but HOW?)",884,5000000,1636329600,70,7,2,75],["2eUWT9cI23o","3D Printed Solids of Constant Width",304,3700000,1510099200,70,4,2,73],["J--QVhGheP4","BEST FREE CAD / 3D Modelling software 2024 - I tested them all!",2498,1900000,1731024000,70,7,2,72],["y44_xMFsPQQ","This Mechanism only turns Clockwise, no matter what.",442,3100000,1541635200,70,7,2,72],["YPAXeBuq9qU","The Ultimate Guide to Perfect 3D Prints",1047,2400000,1667865600,70,7,2,71],["Cuqao90C76U","Top 3 FREE 3D Design Software 2019",415,2400000,1573171200,70,7,2,71],["_QRb54zVPfQ","Stop Ender 3 Stringing with this One Weird Trick! (ok but seriously)",930,2100000,1573171200,70,7,2,70],["3kW9SnK4LKc","3D Print parts TWICE as fast with simple slicer tweaks. 3DP101",882,2000000,1573171200,70,7,2,70],["1xDVbu-WaFo","Hugging Face Journal Club - DeepSeek R1",2606,53000,1762560000,79,3,3,77],["00GKzGyWFEs","Welcome to the Hugging Face course",273,437000,1667865600,79,0,2,77],["F2jd5WuT-zg","Steering LLM Behavior Without Fine-Tuning",1064,145000,1767830400,79,3,2,76],["H39Z_720T5s","The Transformer architecture",165,103000,1636329600,79,0,3,75],["tiZFewofSLM","The pipeline function",276,207000,1636329600,79,0,2,73],["VqyKUuW9V1g","Train AI Robots Without Writing Code! (Introducing LeLab)",293,41000,1780876800,79,7,2,70],["CKFUx7COIwQ","Workshop: Training Agents at the frontier of AI Industry",7095,33000,1793059200,79,2,2,69],["qaP4pymNrEc","RTX 5090 vs M5 Ultra: Which AI Powerhouse Wins?",200,32000,1792281600,79,1,2,69],["MUqNwgPjJvQ","Transformer models: Encoders",286,107000,1636329600,79,0,2,69],["0_4KEb08xrE","Transformer models: Encoder-Decoders",407,118000,1636329600,79,0,2,69],["QgH9sr7G13Q","The most cited paper of the century is a brilliant hack",2119,1800000,1788825600,2,0,3,76],["UZDiGooFs54","The moment we stopped understanding AI [AlexNet]",1058,2500000,1731024000,2,0,3,74],["f8CXG7dS-D0","The most beautiful equation in math, explained visually [Euler’s Formula]",1617,1800000,1731024000,2,4,3,72],["0VLAoVGf_74","How DeepSeek Rewrote the Transformer [MLA]",1089,1000000,1762560000,2,0,3,71],["T647CGsuOVU","Imaginary Numbers Are Real [Part 1: Introduction]",347,8600000,1446940800,2,4,2,71],["D8GOeCFFby4","The most complex model we actually understand",2129,902000,1767830400,2,3,3,70],["l-9ALe3U-Fg","ChatGPT is made from 100 million of these [The Perceptron]",1441,825000,1762560000,2,0,3,69],["NrO20Jb-hy0","The Misconception that Almost Stopped AI [How Models Learn Part 1]",1377,683000,1762560000,2,0,3,68],["5eqRuVp65eY","AI can't cross this line and we don't know why.",1447,2500000,1731024000,2,3,2,67],["nD-Dco7xSSU","Oppenheimer's Apocalypse Math",332,4100000,1699401600,2,4,2,67],["FgakZw6K1QQ","StatQuest: Principal Component Analysis (PCA), Step-by-Step",1318,3700000,1541635200,3,0,3,76],["yIYKR4sgzI8","StatQuest: Logistic Regression",528,2800000,1541635200,3,0,3,74],["4b5d3muPQmA","StatQuest: K-means clustering",510,2200000,1541635200,3,0,3,73],["rzFX5NWojp0","The Normal Distribution, Clearly Explained!!!",313,2100000,1510099200,3,4,3,73],["efR1C6CvhmE","Support Vector Machines Part 1 (of 3): Main Ideas!!!",1232,1800000,1573171200,3,0,3,72],["4jRBRDbJemM","ROC and AUC, Clearly Explained!",977,2000000,1573171200,3,0,3,72],["sDv4f4s2SB8","Gradient Descent, Step-by-Step",1434,1800000,1573171200,3,0,3,72],["EuBBz3bI-aA","Machine Learning Fundamentals: Bias and Variance",396,1800000,1541635200,3,0,3,72],["XepXtl9YKwc","Maximum Likelihood, clearly explained!!!",372,1800000,1510099200,3,4,3,72],["vemZtEM63GY","p-values: What they are and how to interpret them",681,1500000,1604793600,3,4,3,71],["X117w2Rark8","Jev - The Ultimate Classification Model?",979,455000,1791417600,29,1,2,76],["PTuGGdDuyPI","Qwen3.8-27B & How to Serve it Fast",1080,273000,1788825600,29,1,2,73],["53wDOI_7x8I","Open Jev Models Are Here!!",1292,211000,1791417600,29,1,2,72],["Us5ZFp16PaU","Fine-tuning LLMs with PEFT and LoRA",935,194000,1699401600,29,0,3,71],["OrcsMi2-IcM","NEW DeepSeek Harness -  Super Agent!",1073,157000,1793750400,29,2,2,70],["2RA7jWJvQ1U","Which is The Best Qwen3.8-27B?",1279,145000,1791417600,29,1,2,70],["ox1mW2N9Z_Y","MiniCPM5 - Just How Good Can a 1B Model Be?",1254,165000,1783468800,29,1,2,70],["YEZHU4LSUfU","DeepSeek OCR - More than OCR",650,233000,1762560000,29,1,2,70],["vSCYL8ySh9g","Dory - NVIDIA's Looped Transformer",905,138000,1793664000,29,3,2,69],["5XTJRU9na3Y","Gemini 4 Argon",611,127000,1791417600,29,8,2,69],["oJNHXPs0XDk","Neural Network Architectures & Deep Learning",549,874000,1573171200,57,0,3,76],["gXbThCXjZFM","Singular Value Decomposition (SVD): Overview",404,704000,1604793600,57,4,3,75],["lKXW7DRyyro","Div, Grad, and Curl: Vector Calculus Building Blocks for PDEs [Divergence, Gradient, and Curl]",782,579000,1667865600,57,4,3,74],["fkf4IBRSeEc","Principal Component Analysis (PCA)",826,509000,1604793600,57,0,3,73],["nbBvuuNVfco","Singular Value Decomposition (SVD): Mathematical Overview",771,526000,1604793600,57,4,3,73],["sQqniayndb4","Probability and Statistics: Overview",1783,778000,1731024000,57,4,2,72],["0MNVhXEX9to","Reinforcement Learning: Machine Learning Meets Control Theory",1563,428000,1636329600,57,0,3,72],["E8HeD-MUrjY","The Fast Fourier Transform (FFT)",526,436000,1604793600,57,4,3,72],["nl9TZanwbBk","The Discrete Fourier Transform (DFT)",1056,450000,1604793600,57,4,3,72],["-zrY7P2dVC4","Physics Informed Neural Networks (PINNs) [Physics Informed Machine Learning]",2072,235000,1731024000,57,0,3,71],["RyvXxApfHkk","Lesson 3A: What is generative AI? (Deep Dive) | AI Fluency: Framework & Foundations Course",381,814000,1762560000,75,0,3,76],["rKV5JcALQoQ","The different levels of how Claude thinks",327,532000,1783468800,75,3,3,75],["gv0WHhKelSE","Claude Code best practices | Code w/ Claude",1553,594000,1762560000,75,2,3,74],["2YCaBqP8muw","Lesson 7: Effective prompting techniques (Deep Dive) | AI Fluency: Framework & Foundations Course",715,599000,1762560000,75,2,3,74],["T9aRN5JkmL8","AI prompt engineering: A deep dive",4603,612000,1731024000,75,2,3,72],["fGKNUvivvnc","Interpretability: Understanding how AI models think",3543,385000,1762560000,75,3,3,71],["UAmKyyZ-b9E","Introducing Cowork: Claude Code for the rest of your work",69,960000,1770508800,75,8,2,70],["AJpK3YTTKZ4","Introducing Claude Code",235,981000,1762560000,75,2,2,70],["rBJnWMD0Pho","Let Claude handle work in your browser",96,725000,1767830400,75,2,2,69],["0vZ_UVLhSQQ","Getting started with Claude.ai",319,2400000,1765152000,75,2,1,69],["firXjwZ_6KI","Flow-Matching vs Diffusion Models explained side by side",968,61000,1762560000,10,0,3,75],["vrF3MtGwD0Y","MAMBA and State Space Models explained | SSM explained",1347,99000,1731024000,10,0,3,75],["KEv-F5UkhxU","What is LoRA? Low-Rank Adaptation for finetuning LLMs EXPLAINED",502,114000,1699401600,10,0,3,74],["J87hffSMB60","How does Stable Diffusion work? – Latent Diffusion Models EXPLAINED",796,103000,1667865600,10,0,3,73],["SndHALawoag","Swin Transformer paper animated and explained",670,101000,1636329600,10,0,3,73],["344w5h24-h8","Diffusion models explained. How does OpenAI's GLIDE work?",1003,99000,1667865600,10,0,3,72],["1biZfFLPRSY","Positional embeddings in transformers EXPLAINED | Demystifying positional encodings.",580,97000,1636329600,10,0,3,72],["FWFA4DGuzSc","The Transformer neural network architecture EXPLAINED. “Attention is all you need”",608,92000,1604793600,10,0,3,72],["XZLc09hkMwA","Direct Preference Optimization: Your Language Model is Secretly a Reward Model | DPO paper explained",535,44000,1731024000,10,0,3,71],["6BPl81wGGP8","UMAP explained | The best dimensionality reduction?",556,75000,1636329600,10,4,3,71],["kmAc1nDizu0","The Boundary of Computation",779,1300000,1699401600,58,5,3,75],["6htbyY3rH1w","Is the Future of Linear Algebra.. Random?",2111,427000,1731024000,58,4,3,70],["NFo9v_yKQXA","Reinforcement Learning, by the Book",1099,229000,1667865600,58,0,3,63],["UBDgSHPxVME","Gaussian Processes",1427,194000,1636329600,58,0,3,62],["_j6pvGEchWU","Bellman Equations, Dynamic Programming, Generalized Policy Iteration | Reinforcement Learning Part 2",1293,161000,1667865600,58,0,3,61],["-9JM9suCIHs","The Kelly Criterion",541,165000,1667865600,58,4,3,61],["Bs_VjCqyDfU","What Happens When All Training Data is AI Generated?",1094,51000,1762560000,58,3,3,60],["kpOWmwA6tJc","Why Don't AI Agents Work (Yet)?",1065,50000,1762560000,58,2,3,60],["q-Bbxvxpqvw","The Most Important (and Surprising) Result from Information Theory",550,107000,1699401600,58,4,3,59],["jlh21U2texo","What happens at the Boundary of Computation?",899,103000,1699401600,58,5,3,59],["PZRI1IfStY0","Floating Point Numbers - Computerphile",556,2500000,1415404800,59,5,3,75],["8ZtInClXe1Q","How NOT to Store Passwords! - Computerphile",564,2400000,1415404800,59,5,3,75],["1S0aBV-Waeo","Running a Buffer Overflow Attack - Computerphile",1050,2100000,1478563200,59,5,3,74],["MijmeoH9LT4","Characters, Symbols and the Unicode Miracle - Computerphile",577,2100000,1383868800,59,5,3,74],["QRYzre4bf7I","How TOR Works- Computerphile",859,1800000,1510099200,59,5,3,73],["O4xNJsjtN6E","AES Explained (Advanced Encryption Standard) - Computerphile",854,1500000,1604793600,59,5,3,72],["ciNHn38EyRc","Running an SQL Injection Attack - Computerphile",1031,4600000,1478563200,59,5,2,72],["DMtFhACPnTY","SHA: Secure Hashing Algorithm - Computerphile",621,1300000,1510099200,59,5,3,71],["ySN5Wnu88nE","A* (A Star) Search Algorithm - Computerphile",844,1300000,1510099200,59,5,3,71],["GazC3A4OQTE","Dijkstra's Algorithm - Computerphile",643,1400000,1510099200,59,5,3,71],["bOYE6E8JrtU","What is a Transformer? (Transformer Walkthrough Part 1/2)",3781,46000,1699401600,15,0,3,74],["KV5gbOmHbjU","A Walkthrough of A Mathematical Framework for Transformer Circuits",10214,58000,1667865600,15,3,3,74],["Tgq7E4YcPKQ","What Happened With Sparse Autoencoders?",10156,11000,1765152000,15,3,3,70],["dsjUDacBw8o","Implementing GPT-2 From Scratch (Transformer Walkthrough Part 2/2)",4765,24000,1699401600,15,0,3,70],["Aroazwb_QW8","Bitter Lesson-Pilled Interp: A Live Paper Review (Activation Oracles & PCD)",11001,4700,1767830400,15,3,3,65],["s3HgyprprhM","How Reasoning Models Break Mechanistic Interpretability Techniques",2541,4600,1765152000,15,3,3,65],["nCZN09Wjboc","How To Interpret Chain Of Thought: A Walkthrough",10373,4900,1762560000,15,3,3,65],["e0V0pYh4M7M","How Aligned Is Claude? A Live Review of the Opus 4.5 System Card",8590,3800,1767830400,15,3,3,64],["VQ_7zLXHf3s","What do models learn during finetuning? A model diffing paper walkthrough w/ Clement & Julian",10493,4300,1762560000,15,3,3,64],["R3nbXgMnVqQ","A Walkthrough of Toy Models of Superposition w/ Jess Smith",8975,10000,1699401600,15,3,3,64],["opyJVz_T1GY","Nvidia GPU UNLOCK for Local AI - 3090's Prefill 62% FASTER!",344,121000,1793059200,25,1,3,74],["Tq_cmN4j2yY","Deepseek R1 671b Running and Testing on a $2000 Local AI Server",3164,489000,1762560000,25,1,2,73],["3XC8BA5UNBs","AI Home Server 24GB VRAM $750 Budget Build and LLM Benchmarking",1790,267000,1762560000,25,1,2,69],["Oa3MUGsFays","The Only Local AI GPUs Worth Buying Right Now (Ranked)",1228,149000,1791417600,25,1,2,68],["WRi0jApo9NM","BEST Local Ai Motherboard CPU Combo You've Never Heard Of",331,157000,1780876800,25,1,2,68],["yFKOOK6qqT8","Deepseek R1 671b Running LOCAL AI LLM is a ChatGPT Killer!",1153,656000,1762560000,25,1,1,68],["WOlXMUKJeuI","Testing Strata Local AI on $769 Budget AI Rig",616,121000,1792281600,25,1,2,67],["x7s6Mf4xF9A","Running Qwen 3.8 Flash Next on a $1000 Local AI Build with Strata",624,126000,1791676800,25,1,2,67],["975ILFNTKfk","Is Bonsai 2 27B REALLY 98% of Qwen 3.8 27B fp16? Local AI Test",941,127000,1791417600,25,1,2,67],["So7tqRSZ0s8","ULTIMATE Local AI Quad 3090 Build",1790,165000,1762560000,25,1,2,67],["LnzuMJLZRdU","“Hello, world” from scratch on a 6502 — Part 1",1645,5800000,1573171200,61,5,3,74],["wdgULBpRoXk","How does a USB keyboard work?",2055,3700000,1636329600,61,5,3,72],["4alV5LzHLE4","Let's build a voltage multiplier!",992,3300000,1699401600,61,7,3,71],["yl8vPW5hydQ","How do CPUs read machine code? — 6502 part 2",2953,3200000,1573171200,61,5,3,71],["sTu3LwpF6XI","Making logic gates from transistors",782,3100000,1446940800,61,5,3,70],["l7rce6IQDWs","The world's worst video card?",1967,7100000,1573171200,61,7,2,69],["eq5YpKHXJDM","Reliable data transmission",2634,1800000,1541635200,61,5,3,67],["yOyaJXpAYZQ","Comparing C to machine language",602,5800000,1446940800,61,5,2,67],["a6EWIh2D1NQ","Hacking a weird TV censoring device",1259,4400000,1699401600,61,7,2,66],["dXdoim96v5A","8-bit CPU control logic: Part 1",1672,1400000,1510099200,61,5,3,66],["B2A2pGrDG8I","What Is the Pigeonhole Principle?",503,3500000,1604793600,64,4,3,74],["K1a2Bk8NrYQ","Understanding B-Trees: The Data Structure Behind Modern Databases",759,1200000,1731024000,64,5,3,71],["EFg3u_E6eHU","How Dijkstra's Algorithm Works",511,2100000,1604793600,64,5,3,71],["INEtYZqtjTo","Understanding Logic Gates",448,1400000,1604793600,64,5,3,69],["85oMrKd8afY","Diffie-Hellman Key Exchange: How to Share a Secret",549,476000,1731024000,64,5,3,65],["bbkcEiUjehk","How Floating-Point Numbers Are Represented",596,176000,1762560000,64,5,3,62],["C4ATDMIz5wc","AES: How to Design Secure Encryption",937,314000,1699401600,64,5,3,60],["SLgZhpDsrfc","Minimax: How Computers Play Games",877,276000,1699401600,64,5,3,59],["w6E7aQnA4Ws","A Computer Built With Dominos",490,868000,1604793600,64,5,2,59],["AE5I0xACpZs","What Is a Binary Heap?",525,290000,1604793600,64,5,3,59],["bCz4OMemCcA","Attention is all you need (Transformer) - Model explanation (including math), Inference and Training",3484,804000,1699401600,4,0,3,73],["ISNdQcPhsts","Coding a Transformer from scratch on PyTorch, with full explanation, training and inference.",10764,409000,1699401600,4,0,3,69],["zy8ChVd_oTM","Flash Attention derived and coded from first principles with Triton (Python)",27498,109000,1762560000,4,6,3,66],["vAmKB7iPkWw","Coding a Multimodal (Vision) Language Model from scratch in PyTorch with full explanation",20765,147000,1731024000,4,0,3,65],["ZBKpAp_6TGI","Coding Stable Diffusion from scratch in PyTorch",18212,225000,1699401600,4,0,3,65],["XoGvCBRnwLs","Building a distributed training framework from first principles",70485,50000,1788825600,4,6,3,63],["rhZgXNdhWDY","Retrieval Augmented Generation (RAG) Explained: Embedding, Sentence BERT, Vector Database (HNSW)",2964,92000,1731024000,4,2,3,62],["Mn_9W1nCFLo","LLaMA explained: KV-Cache, Rotary Positional Embedding, RMS Norm, Grouped Query Attention, SwiGLU",4255,131000,1699401600,4,0,3,62],["qGyFrqc34yc","Reinforcement Learning from Human Feedback explained with math derivations and the PyTorch code.",8113,79000,1731024000,4,0,3,61],["8Q_tqwpTpVU","Mamba and S4 Explained: Architecture, Parallel Scan, Kernel Fusion, Recurrent, Convolution, Math",4469,68000,1731024000,4,0,3,61],["l-OLgbdZ3kk","The Brain’s Learning Algorithm Isn’t Backpropagation",1612,775000,1762560000,16,3,3,73],["SmZmBKc7Lrs","The Most Important Algorithm in Machine Learning",2408,1200000,1731024000,16,0,3,73],["ceFFEmkxTLg","How Your Brain Chooses What to Remember",1039,1200000,1762560000,16,3,2,69],["jnxqHcObNK4","Wavelets: a mathematical microscope",2069,801000,1667865600,16,4,3,68],["KHVR587oW8I","The Key Equation Behind Probability",1584,425000,1731024000,16,4,3,67],["_bqa_I5hNAo","Generative Model That Won 2024 Nobel Prize",1984,384000,1731024000,16,0,3,66],["q7seckj1hwM","What Textbooks Don't Tell You About Curve Fitting",1094,198000,1762560000,16,4,3,65],["zOmhHE2xctw","The Core Equation Of Neuroscience",1395,586000,1731024000,16,4,2,62],["vTTlzmCRwU4","Differential Equations: The Language of Change",1404,189000,1731024000,16,4,3,62],["hmtQPrH-gC4","Dendrites: Why Biological Neurons Are Deep Neural Networks",1528,281000,1699401600,16,3,3,62],["-PjMC0gyH9s","NVIDIA Jetson Orin Nano Super COMPLETE Setup Guide & Tutorial",3395,354000,1762560000,28,7,2,73],["9Z9rPZavjUU","Claude Fable 5.1 Is INSANE – Hands-On With the BEST Model Yet!",2383,184000,1788825600,28,8,2,72],["mfsUVN8ykLA","Gemini 4 Argon Is INSANE – Hands-On With Google’s New Model!",2230,149000,1792281600,28,8,2,71],["ux6Lafw7en0","Claude Opus 5.5 Is INSANE – Hands-On With the BEST Model Yet!",2260,153000,1791417600,28,8,2,71],["SX5f8IIBriU","Liquid LFM3 Is INSANE – Tiny Local AI That ANYONE Can Run!",1592,142000,1793232000,28,1,2,70],["dk1OY5z08_Y","The New DeepSeek V4.1 Flash Is INSANE – The BEST Open Model Yet?",2134,137000,1792886400,28,8,2,70],["-56F0u0xN1A","Bonsai 27B Deep Dive – 1-Bit, Ternary & Full Precision Compared!",2025,44000,1786147200,28,1,3,70],["4MRMLerpcQc","Kimi K3.1 Pro Is HERE – Does THIS Really Beat Claude & GPT?",2738,120000,1792886400,28,8,2,69],["OO_wTt_ogtU","Ornith 1.5 35B is VERY Good - Full Test & Quant Comparison!",1900,36000,1788825600,28,1,3,69],["OA5cICIzD-c","Bonsai 2 27B First Test – Is THIS the BEST Single-GPU AI Model?",1714,92000,1791417600,28,1,2,68],["d3qGQ2utl2A","Day 1 of Learn Fusion 360 in 30 Days for Complete Beginners! - 2023 EDITION",787,3300000,1667865600,72,7,2,73],["qvrHuaHhqHI","Fusion 360 Tutorial for Absolute Beginners (2020)",2072,1800000,1573171200,72,7,2,69],["4G2E_DqQteM","Day 1 of Learn Autodesk Fusion in 30 Days! (2026 EDITION)",1172,527000,1775606400,72,7,2,68],["DfAfxae8aRc","Day 2 of Learn Fusion 360 in 30 Days for Complete Beginners! - 2023 EDITION",715,1200000,1667865600,72,7,2,67],["CeMHqa9Pxn8","Import and Edit STL Files in Fusion 360 2024",859,1200000,1667865600,72,7,2,67],["lQ-MYnyxh7M","CNC Basics - What You Need To Get Started",558,1300000,1604793600,72,7,2,67],["TzG2deElWqI","Bodies vs Components | Fusion 360 Core Concept",927,360000,1604793600,72,7,3,66],["zmGbUH-P1lE","Don't open Fusion 360 until you learn these 8 terms",358,298000,1778198400,72,7,2,65],["_x2Q44UTyrA","Day 2 of Learn Autodesk Fusion in 30 Days! (2026 EDITION)",1015,256000,1775606400,72,7,2,64],["kj3_doWX2Kg","Fusion 360 Components and Assemblies Explained | Day 13 of Learn Fusion 360 in 30 Days",682,237000,1699401600,72,7,3,64],["amEUIuBKwvg","A Complete Guide to Claude Code - Here are ALL the Best Strategies",3017,151000,1762560000,20,2,3,72],["V_0dNE-H2gw","Run ALL Your AI Locally in Minutes (LLMs, RAG, and more)",1219,595000,1731024000,20,1,2,71],["Egeuql3Lrzg","Context Engineering is the New Vibe Coding (Learn this Now)",1376,325000,1762560000,20,2,2,70],["mNcXue7X8H0","The Ultimate Guide to Local AI and AI Agents (The Future is Here)",9517,283000,1762560000,20,1,2,69],["fg0_0M8kZ8g","Turn ANY File into LLM Knowledge in SECONDS",1281,260000,1762560000,20,2,2,68],["zbmuiaPuiNM","Google Just Dropped a Masterclass on Agentic Engineering (It's SO Good)",1316,168000,1783468800,20,2,2,67],["-1K_ZWDKpU0","Claude Code's Agent Teams Are Insane - Multiple AI Agents Coding Together in Real Time",1216,187000,1772928000,20,2,2,67],["7huCP6RkcY4","I Built Self-Evolving Claude Code Memory w/ Karpathy's LLM Knowledge Bases",1164,161000,1775606400,20,2,2,66],["1FiER-40zng","Full Guide - Build Your Own AI Second Brain with Claude Code",1504,146000,1775606400,20,2,2,66],["p0FERNkpyHE","Introducing RAG 2.0: Agentic RAG + Knowledge Graphs (FREE Template)",2328,172000,1762560000,20,2,2,66],["d5x0JCZbAJs","From 0 to Production - The Modern React Tutorial (RSCs, Next.js, Shadui, Drizzle, TS and more)",10992,583000,1731024000,21,5,3,72],["I82j7AzMU80","How does Claude Code *actually* work?",2363,240000,1778198400,21,2,3,70],["jC_z1vL1OCI","I finally know how CPUs work (w/ Casey Muratori)",12001,313000,1762560000,21,6,3,70],["xq0eB8vxeqk","Agents Sucked At UI (Until Today)",1905,382000,1793836800,21,2,2,67],["ejjBbaq9RmY","Getting the most out of Opus 5.5",1722,244000,1791417600,21,2,2,64],["GcNu6wrLTJc","Delete your CLAUDE.md (and your AGENT.md too)",1756,304000,1772928000,21,2,2,64],["c-hKSbzooAg","My Biggest Tutorial Ever (Build A FULL Google Drive Clone with React, Next, TypeScript and more)",12454,277000,1762560000,21,5,2,63],["INLq9RPAYUw","Why is every React site so slow?",832,301000,1762560000,21,5,2,63],["0oXOOlqVu5M","So I tried Matt's skills...",2302,532000,1788825600,21,2,1,62],["JMYspR42HFM","Claude Code vs Codex vs Cursor (an honest comparison)",2277,178000,1780876800,21,2,2,62],["l30sADfDiM8","The ULTIMATE Raspberry Pi 5 NAS",1934,4400000,1731024000,27,5,2,72],["KUUuxkICOLw","Qualcomm made a Pi 5 killer. Good luck getting one.",548,1500000,1791676800,27,6,2,70],["o1sN1lB76EA","OpenAI's nightmare: Deepseek R1 on a Raspberry Pi",258,2300000,1762560000,27,1,2,70],["UEtpaiODNs0","Cute, but powerful: meet NanoCluster, a tiny supercomputer",1356,1500000,1762560000,27,6,2,68],["A2B7oI0FYqo","Intel just matched Apple Silicon. Seriously.",347,1000000,1786147200,27,6,2,67],["eb48MdtNaDQ","Bambu Lab 3D printers: Never again",211,1000000,1780876800,27,7,2,67],["3hFas54xFtg","I replaced my Apple TV—with a Raspberry Pi",972,2000000,1731024000,27,7,2,67],["4l9xU3_36Xs","Formlabs finally made a cheaper printer",1418,733000,1792281600,27,7,2,66],["ltofM26akZQ","Raspberry Pi locks down RAM upgrades in Firmware",206,748000,1791417600,27,6,2,66],["UtLyX72-688","Raspberry Pi 5: Getting Started",1483,1600000,1731024000,27,7,2,66],["SOQ6F7HMfSc","The Rise and Fall of the Cray Supercomputer",1949,1000000,1731024000,56,6,3,72],["dnHdqPBrtH8","Why the Soviet Computer Failed",1137,4100000,1667865600,56,5,2,71],["47ovT_hYlt8","The Tragedy of Compaq",1994,1600000,1762560000,56,6,2,70],["rSpzAVpnXo4","What Happened to the Capacitors in 2002?",1338,1300000,1762560000,56,6,2,69],["1fOA85xtYxs","How an ASML Lithography Machine Moves a Wafer",975,647000,1731024000,56,6,3,69],["Pzy_KOBddRA","The First Transistors",1490,478000,1731024000,56,5,3,68],["5RPFfPtgw7g","The Gate-All-Around Transistor is Coming",944,542000,1731024000,56,6,3,68],["sV7C6Ezl35A","A Cult AI Computer’s Boom and Bust",2721,282000,1762560000,56,6,3,67],["1lG7lFLXBIs","The Evolution of the Operating System",1647,433000,1731024000,56,5,3,67],["XHrQ-Pmvwao","TSMC's First Breakthrough: The Copper/Low-K Interconnect Transition",2090,660000,1699401600,56,6,3,67],["DdrrrM_IjnA","3D printer nozzles are getting WILD",1189,1000000,1791417600,69,7,2,72],["ikjOEaPY6Rk","I bought the cheapest 3D printer on AliExpress!",1362,2200000,1731024000,69,7,2,72],["8_adY2K-YIc","Things you should know about PETG",466,1200000,1541635200,69,7,3,72],["HgEEtk85rAY","A better way to add threads to your 3D prints",687,1100000,1762560000,69,7,2,70],["nb-Bzf4nQdE","3D Printing Basics! (Ep1)",681,900000,1604793600,69,7,3,70],["il9bNWn66BY","Autoleveling on 3D printers: 9 myths and 12 sensors tested!",788,794000,1510099200,69,7,3,70],["BR_yko0gr-Y","Making a Digital Dashboard! (w/ Google Calendar integration)",1166,2100000,1573171200,69,7,2,69],["unmvWkfpmRE","Things you should know about PLA",465,734000,1541635200,69,7,3,69],["6TZ2R88rPoo","3D Printing Basics: Resin Printers! (Ep3)",776,496000,1604793600,69,7,3,67],["RbtINaQOzDE","3D Printing Basics: Choosing a printer! (Ep2)",1245,507000,1604793600,69,7,3,67],["aircAruvnKk","But what is a neural network? | Deep learning chapter 1",1120,25000000,1510099200,1,0,3,71],["LPZh9BOjkQs","Large Language Models explained briefly",478,8400000,1762560000,1,0,3,70],["wjZofJX0v4M","Transformers, the tech behind LLMs | Deep Learning Chapter 5",1634,11000000,1731024000,1,0,3,69],["r6sGWTCMz2k","But what is a Fourier series?  From heat flow to drawing with circles | DE4",1487,19000000,1573171200,1,4,3,69],["spUNpyF58BY","But what is the Fourier Transform?  A visual introduction.",1182,12000000,1541635200,1,4,3,67],["fNk_zzaMoSs","Vectors | Chapter 1, Essence of linear algebra",592,13000000,1478563200,1,4,3,67],["RQWpF2Gb-gU","But what is quantum computing?  (Grover's Algorithm)",2214,4200000,1762560000,1,5,3,66],["WUvTyaaNkzM","The essence of calculus",1025,11000000,1510099200,1,4,3,66],["IHZwWFHWa-w","Gradient descent, how neural networks learn | Deep Learning Chapter 2",1233,9700000,1510099200,1,0,3,65],["eMlx5fFNoYc","Attention in transformers, step-by-step | Deep Learning Chapter 6",1570,4800000,1731024000,1,0,3,64],["00Y-p62sk0s","My M5 Max, Gemma 4, MLX LOCAL Stack. (This KILLS MODEL PROVIDERS)",2344,190000,1778198400,19,1,2,71],["DNDYYHsypZg","Pi Durable PROACTIVE Agents: Jev Triggers My ON-CALL Pi Agent",2116,79000,1792281600,19,2,2,67],["haUfb1ievTE","My Super Simple Software Factory (For Agentic Engineers)",1793,84000,1786147200,19,2,2,67],["J5B9UGTuNoM","I’m HOOKED on Claude Code Hooks: Advanced Agentic Coding",1851,125000,1762560000,19,2,2,67],["hGg3nWp7afg","Agentic Claude Code: 3 Codebase Folders for TOP 1% AI Coding",1471,112000,1762560000,19,2,2,67],["_U-O5lYhJ7Q","10 Levels of Jev For Agentic Engineers",2118,209000,1791417600,19,2,1,66],["f8cfH5XX-XU","The Pi Coding Agent: The ONLY REAL Claude Code COMPETITOR",3097,286000,1772928000,19,2,1,66],["7B2HJr0Y68g","My Claude Code Sub Agents BUILD THEMSELVES",1838,103000,1762560000,19,2,2,66],["PIdETjcXNIk","Pi to Pi: Two-Way Agent Orchestration with the Pi Coding Agent",2092,69000,1780876800,19,2,2,65],["kFpLzCVLA20","I finally CRACKED Claude Agent Skills (Breakdown For Engineers)",1634,84000,1762560000,19,2,2,65],["u8dIpwd6tzo","Everyone gets this Wrong when 3D Printing Carbon Fiber Nylon",1315,851000,1762560000,67,7,3,69],["BT04glGDjB4","Recycling Failed 3D Prints with a DIY Filament Extruder: Artme3D",1210,3800000,1731024000,67,7,2,68],["O6d1RKYapFI","Turning 3D Print Waste into New Filament!",768,4000000,1699401600,67,7,2,66],["vqWwUx8l_Io","Recycle your failed 3D prints! Make new filament at home.",543,4500000,1573171200,67,7,2,66],["0Hi9lBjZGIM","Anti-Aliasing for FDM 3D Printing Is Finally Here (Micro-Non-Planar Printing)",995,399000,1770508800,67,7,3,65],["7aJ4dDyyf20","Stop Printing Flat: The 45° Secret for Stronger Parts",1025,1500000,1762560000,67,7,2,65],["fbSQvJJjw2Q","Which LAYER HEIGHT gives you the STRONGEST 3D prints?",813,3700000,1573171200,67,7,2,65],["B5g7R53hcH4","Are Our Heat-Set Insert Datasheets Wrong?",1109,279000,1775606400,67,7,3,63],["ycGDR752fT0","The BEST 3D printing material? Comparing PLA, PETG & ASA (ABS) - feat. PRUSAMENT by Josef Prusa",1033,2700000,1604793600,67,7,2,63],["upELI0HmzHc","TESTING 3D printed INFILL PATTERNS for their STRENGTH",515,809000,1541635200,67,7,3,63],["XJC5WB2Bwrc","Why Some Projects Use Multiple Programming Languages",1172,1500000,1762560000,62,5,2,68],["HjneAhCy2N4","HOW TRANSISTORS RUN CODE?",868,1600000,1731024000,62,5,2,66],["Y2mDwW2pMv4","IPC: To Share Memory Or To Send Messages",855,295000,1762560000,62,5,3,65],["tadUeiNe5-g","How CPUs Interact with So Many Different Devices",1408,815000,1762560000,62,5,2,64],["GYlNoAMBY6o","CRAFTING A CPU TO RUN PROGRAMS",1189,394000,1731024000,62,5,3,64],["SwIPOf2YAgI","The Weird Way Linux Creates Processes",1071,183000,1770508800,62,5,3,63],["M9HHWFp84f0","Why Are Threads Needed On Single Core Processors",967,658000,1762560000,62,5,2,63],["QAzuAn3nFGo","The 80’s Algorithm to Avoid Race Conditions (and Why It Failed)",1145,188000,1762560000,62,5,3,62],["rM9BjciBLmg","HOW TRANSISTORS REMEMBER DATA",1018,817000,1731024000,62,5,2,62],["8gnvx11j288","Who Actually Calls the main Function?",587,111000,1793404800,62,5,3,61],["h7apO7q16V0","The Fast Fourier Transform (FFT): Most Ingenious Algorithm Ever?",1703,2200000,1636329600,63,4,3,68],["0me3guauqOU","The Unreasonable Effectiveness of JPEG: A Signal Processing Approach",2088,1200000,1667865600,63,5,3,64],["EFUYNoFRHQI","How PNG Works: Compromising Speed for Quality",1920,752000,1667865600,63,5,3,61],["6oMZb3yP_H8","How Computers Draw Weird Shapes (Marching Squares)",1680,431000,1636329600,63,5,3,58],["aPQY__2H3tE","5 Simple Steps for Solving Dynamic Programming Problems",1287,1500000,1604793600,63,5,2,58],["ngCos392W4w","5 Simple Steps for Solving Any Recursive Problem",1263,1500000,1604793600,63,5,2,58],["GiDsjIBOVoA","The Traveling Salesman Problem: When Good Enough Beats Perfect",1827,380000,1667865600,63,5,3,57],["ajv46BSqcK4","A Strange But Elegant Approach to a Surprisingly Hard Problem (GJK Algorithm)",1913,381000,1636329600,63,4,3,57],["B3y0RsVCyrw","Huffman Codes: An Information Theory Perspective",1751,293000,1636329600,63,5,3,55],["LFKZLXVO-Dg","Introduction to Graph Theory: A Computer Science Perspective",986,806000,1604793600,63,5,2,55],["FluKUJyeYD8","Did OpenAI actually build AGI? GPT-6 Astra first look",447,4200000,1788825600,65,8,2,68],["wajoP_5Cpw8","Every app on Earth is now open source... whether it likes it or not",356,6700000,1792281600,65,5,1,64],["MtxP2pyCvYA","Every operating system concept in one video…",691,795000,1778198400,65,5,3,64],["hdHjjBS4cs8","Brainf**k in 100 Seconds",130,7900000,1667865600,65,5,2,64],["WrCjAAl9okA","A $6.3 billion open-weight model just got embarrassed by the French...",370,1600000,1791417600,65,1,2,63],["6fnmXX8RK0s","5 wild data structures every developer should know",270,695000,1762560000,65,5,3,62],["l3L_lETiSzc","The runtime wars are over... and Node won",340,1200000,1792281600,65,5,2,61],["Y5rSSvXfL4g","5 open source tools that replaced my $320/mo AI stack...",326,1300000,1788825600,65,2,2,61],["UuqSy1jPSUw","The most expensive software bug in history...",319,1300000,1788825600,65,5,2,61],["Iq_r7IcNmUk","25 crazy software bugs explained",1010,2600000,1731024000,65,5,2,61],["vSwumoSlZTo","I Tested (Almost) EVERY FILAMENT on Amazon: Every Single Filament Part I",2536,3000000,1636329600,71,7,2,64],["OcMvTfUfNXo","You can copy and print perfect spare parts! (Legally) (Probably)",2056,893000,1775606400,71,7,2,63],["GoEquOdUlGY","Four Million Tiny Mirrors: The Insane Engineering of DLP and the Future of 3D Printing",1081,581000,1667865600,71,7,3,61],["ra_9zU-mnl8","Gridfinity: Your Ultimate Modular Workshop is FREE!",1252,1700000,1667865600,71,7,2,60],["KgwFTv9J7ik","How MakerBot sold us literal hot garbage",1999,544000,1762560000,71,7,2,59],["vBdqbnfvj_A","These 3D filaments SHOULD NOT EXIST! Every 3D Filament Part 7",3040,326000,1788825600,71,7,2,58],["weeG9yOp3i4","The 3D Filament Tier List! Which Should YOU Use?",2674,2300000,1731024000,71,7,1,58],["MqirgLJnmZM","Five 3D-Printing Mistakes You're Making RIGHT NOW!!",1313,1100000,1699401600,71,7,2,58],["mPh1b2EOOz0","The Most Advanced Filaments You Can Print: Every Single Filament Part II",2081,1100000,1636329600,71,7,2,58],["ofI2OCq_QqA","Announcing GRIDFINITY 2.0! Print it now, FOR FREE!",871,948000,1699401600,71,7,2,57],["_7vPNcnYWQ4","Making an unpickable lock. Calling locksmiths",790,18000000,1636329600,66,7,2,55],["aXfTgCCsRSg","What if we made a camera that sees in reverse?",1383,5900000,1762560000,66,7,2,54],["2A2NY29iQdI","TWO Unpickable (?) Locks for Lock Picking Lawyer!",880,14000000,1636329600,66,7,2,54],["cQO2XTP7QDw","I sent robot forgeries to a handwriting expert",1419,11000000,1699401600,66,7,2,53],["X9zXcnSXNF0","Making a robot to carve photos into pumpkins",966,9700000,1604793600,66,7,2,52],["myO8fxhDRW0","Moving hoop won't let you miss",823,28000000,1604793600,66,7,1,51],["vJ43DjwLPGA","Why is this task so difficult for machines?",1627,3300000,1762560000,66,7,2,50],["Gu_1S77XkiM","A simple human task that's insanely hard for a robot",1204,7100000,1667865600,66,7,2,50],["osUTMnDFV30","My GIANT wall painting robot",1404,7800000,1636329600,66,7,2,50],["vsTTXYxydOE","Automatic pool stick vs. strangers",1279,23000000,1636329600,66,7,1,50],["T72nqdC92PM","From Scratch to SOTA: Training a 3B State-Space Vision Model — Krishna Prasad Srinivasan, Sarvam",1269,121000,1791417600,18,0,3,83],["fLUtUkqYHnQ","Everything I Learned Training Frontier Small Models — Maxime Labonne, Liquid AI",1213,150000,1778198400,18,0,3,83],["kPL-6-9MVyA","RAG Agents in Prod: 10 Lessons We Learned — Douwe Kiela, creator of RAG",1016,179000,1762560000,18,2,3,83],["TqC1qOfiVcQ","Claude Agent SDK [Full Workshop] — Thariq Shihipar, Anthropic",6745,157000,1767830400,18,2,3,82],["OkEGJ5G3foU","[Full Workshop] Reinforcement Learning, Kernels, Reasoning, Quantization & Agents — Daniel Han",9748,171000,1762560000,18,0,3,82],["YpY83-kA7Bo","2025 in LLMs so far, illustrated by Pelicans on Bicycles — Simon Willison",1110,160000,1762560000,18,8,3,82],["mR-WAvEPRwE","Anthropic Workshop: Build Agents That Run for Hours — Ash Prabaker & Andrew Wilson",4540,106000,1780876800,18,2,3,81],["Lue8K2jqfKk","Claude Code & the evolution of agentic coding — Boris Cherny, Anthropic",1092,137000,1762560000,18,2,3,81],["RFKCzGlAU6Q","How Claude Code Works - Jared Zoneraich, PromptLayer",3943,102000,1767830400,18,2,3,80],["640KMYtxCeI","Building Multimodal AI Agents From Scratch — Apoorva Joshi, MongoDB",2218,120000,1762560000,18,2,3,80],["JIsgyk0Paic","Reinforcement Learning for Agents - Will Brown, ML Researcher at Morgan Stanley",1097,120000,1762560000,18,0,3,80],["shRR1e2HXMk","How Codex Works — Dominik Kundel, OpenAI",1255,72000,1788825600,18,2,3,79],["04_gN-C9IAo","The Shape of Compute (Chris Lattner of Modular)",4698,51000,1762560000,17,6,3,78],["2czYyrTzILg","From Chaos to Choreography: Multi-Agent Orchestration Patterns That Actually Work — Sandipan Bhaumik",1589,63000,1778198400,18,2,3,78],["fKMB5UlVY1E","Stanford CS25: V4 I Overview of Transformers",4649,76000,1731024000,7,0,3,77],["_IlTcWciEC4","Context Engineering for Agents - Lance Martin, LangChain",3798,45000,1762560000,17,2,3,77],["UsB70Tf5zcE","Training an LLM from Scratch, Locally — Angelos Perivolaropoulos, ElevenLabs",4886,59000,1778198400,18,0,3,77],["bkVCAk9Nsss","Stanford CS330 Deep Multi-Task & Meta Learning - What is multi-task learning? I 2022 I Lecture 1",4318,97000,1699401600,7,0,3,76],["XL07WEc2TRI","Stanford Seminar - Information Theory of Deep Learning, Naftali Tishby",5084,99000,1541635200,7,0,3,76],["MhHEGMFCEB0","OpenAI Codex Masterclass  — Vaibhav Srivastav & Katia Gil Guzman",3719,49000,1778198400,18,2,3,76],["_A367W_qvc8","Gemma 4 Deep Dive — Cassidy Hardin, Researcher, Google DeepMind",1143,45000,1778198400,18,0,3,76],["-cKUW6n8hBU","DSPy: The End of Prompt Engineering - Kevin Madura, AlixPartners",4393,51000,1770508800,18,2,3,76],["PUBc2G0Tj5E","Meet the New Codex CLI",195,243000,1791417600,77,2,3,76],["6nJZopACRuQ","Pre-Training GPT-4.5",2794,350000,1762560000,77,0,3,76],["1DtMiRKg-cs","Understanding the inner thoughts of AI",3186,242000,1786147200,78,3,3,76],["lfHX2hHRMVQ","RL Course by David Silver - Lecture 2: Markov Decision Process",6125,775000,1446940800,78,0,3,76],["25FsKN0f8gQ","[NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton",1699,29000,1767830400,17,3,3,75],["ys6YQ2yfBpM","Your Old Prompts Are Making Claude Worse - Lance Martin, Anthropic",3843,32000,1794009600,18,2,3,75],["gv4w23xYj-0","How to Write Your First AI Eval - SallyAnn DeLucia, Arize AI",2120,34000,1793059200,18,2,3,75],["xzXdLRUyjUg","Lets Build An Agent from Scratch — Kam Lasater",1169,47000,1762560000,18,2,3,75],["cdflu9ZXZGE","Terence Tao on How AI Is Changing Mathematics",76,232000,1780876800,77,4,3,75],["7R52wiUgxZI","DeepMind x UCL | Deep Learning Lectures | 1/12 | Intro to Machine Learning & AI",5118,671000,1604793600,78,0,3,75],["mYSRn6PC1mc","Full Workshop: Build Your Own Deep Research Agents - Louis-François Bouchard, Paul Iusztin, Samridhi",7023,33000,1778198400,18,2,3,74],["-rsTkYgnNzM","Rethinking how we Scaffold AI Agents - Rahul Sengottuvelu, Ramp",992,40000,1762560000,18,2,3,74],["5PFXJYme4ik","Build an LLM from Scratch 6: Finetuning for Classification",8129,32000,1762560000,5,0,3,73],["pvcEQfcO3pk","Statistical Learning: 2.3 Model Selection and Bias Variance Tradeoff",605,61000,1667865600,7,0,3,73],["T3SS931wU0I","Agentic Search vs Vector Search for Coding Agents: We Ran the Eval — Braintrust",1095,24000,1791417600,18,2,3,73],["Jty4s9-Jb78","Jack Morris: Stuffing Context is not Memory, Updating Weights is",3764,31000,1767830400,18,3,3,73],["bk0TmxoZlUY","Evals 101 — Doug Guthrie, Braintrust",2911,34000,1762560000,18,2,3,73],["KUEmEb71vzQ","Function Calling is All You Need — Full Workshop, with Ilan Bigio of OpenAI",6174,34000,1762560000,18,2,3,73],["kDlqpN1JyIw","Vercel AI SDK Masterclass: From Fundamentals to Deep Research",3592,39000,1762560000,18,2,3,73],["pRM_P6UfdIc","Low Level Technicals of LLMs: Daniel Han",10346,57000,1731024000,18,0,3,73],["Q78wvrQ9xsU","Coding Adventure: Simulating Smoke",2500,678000,1762560000,60,5,3,73],["yyJ-hdISgnw","Coding Adventure: Software Rasterizer",3009,641000,1762560000,60,5,3,73],["Qz0KTGYJtUk","Coding Adventure: Ray Tracing",2278,1500000,1699401600,60,5,3,73],["bqtqltqcQhw","Coding Adventure: Boids",515,1700000,1573171200,60,5,3,73],["-MPGU2a67Ls","Meet GPT-5.6",119,149000,1786147200,77,8,3,73],["x4O8pojMF0w","Solving Rubik’s Cube with a Robot Hand",171,532000,1573171200,77,3,3,73],["Nd1-UUMVfz4","RL Course by David Silver - Lecture 3: Planning by Dynamic Programming",5949,477000,1446940800,78,0,3,73],["237Hf7Q3lgg","Build A Reasoning Model From Scratch 6: Reinforcement Learning 1 (Implementing GRPO for RLVR)",5214,20000,1791417600,5,0,3,72],["JQJ_8_jSAoY","Build A Reasoning Model From Scratch 3: The Verifier for Evaluation and RL with Verifiable Rewards",5207,20000,1791417600,5,0,3,72],["BJua0yjO5dk","Build A Reasoning Model From Scratch 2: Loading a Base Model, Text Generation, and KV Caching",5802,22000,1788825600,5,0,3,72],["AdLgPmcrXwQ","Stanford CS25: V4 I Aligning Open Language Models",4581,30000,1731024000,7,0,3,72],["_cR9zDlvP88","Stanford Lecture: Don Knuth—\"Dancing Links\" (2018)",5425,48000,1573171200,7,5,3,72],["pFsfax19yOM","Full Walkthrough: Writing & Using Skills — Nick Nisi and Zack Proser",4863,25000,1778198400,18,2,3,72],["-mRi-B3t6fA","Llamafile: bringing AI to the masses with fast CPU inference: Stephen Hood and Justine Tunney",1045,49000,1731024000,18,1,3,72],["4QOcCGI6xOU","Coding Adventure: Clouds",770,1300000,1573171200,60,5,3,72],["l-hh51ncgDI","Algorithms Explained – minimax and alpha-beta pruning",661,1400000,1541635200,60,5,3,72],["ELh8R7bGlxE","Sol, Terra, and Luna, our GPT‑5.6 family of models are here.",207,143000,1786147200,77,8,3,72],["TCCjZe0y4Qc","DeepMind x UCL RL Lecture Series - Introduction to Reinforcement Learning [1/13]",5392,405000,1636329600,78,0,3,72],["PnHCvfgC_ZA","RL Course by David Silver - Lecture 4: Model-Free Prediction",5821,399000,1446940800,78,0,3,72],["fSytzGwwBVw","Machine Learning Fundamentals: Cross Validation",365,1500000,1541635200,3,0,3,71],["pYxNSUDSFH4","In Statistics, Probability is not Likelihood.",301,1500000,1541635200,3,4,3,71],["J4Wdy0Wc_xQ","StatQuest: Random Forests Part 1 - Building, Using and Evaluating",594,1500000,1541635200,3,0,3,71],["nk2CQITm_eo","Linear Regression, Clearly Explained!!!",1647,1500000,1510099200,3,0,3,71],["msp69FR5NZs","Build A Reasoning Model From Scratch 8: Distillation",4878,18000,1792281600,5,0,3,71],["U8J32Z3qV8s","Stanford CS25: V1 I Mixture of Experts (MoE) paradigm and the Switch Transformer",3944,42000,1667865600,7,0,3,71],["3DKo219ZHMw","Stanford Lecture: Don Knuth - \"Pi and The Art of Computer Programming\" (2019)",3746,43000,1604793600,7,5,3,71],["DVoHvmww2lQ","An image is worth 16x16 words: ViT | Vision Transformer explained",326,83000,1604793600,10,0,3,71],["jrf76uNs77k","The Unreasonable Effectiveness of Reasoning Distillation: using DeepSeek R1 to beat OpenAI o1",1415,15000,1762560000,17,0,3,71],["Uv0nAf_7tyU","Decoupling the Brain from the Hands: Claude Managed Agents (Part 1 of 4) — Anthropic",1060,18000,1792281600,18,2,3,71],["K0X9QDRkIdg","Claude Managed Agents and Evolution of Agentic Surfaces — Gagan Bhat & Isabella Kai He, Anthropic",1884,18000,1788825600,18,2,3,71],["9QebvrrY3KY","Claude for Long-Horizon Tasks — Lance Martin, Anthropic",1519,20000,1786147200,18,2,3,71],["9iN-cPnp7xg","[Evals Workshop] Mastering AI Evaluation: From Playground to Production",5108,27000,1762560000,18,2,3,71],["7UvtU75NXTg","The Laplace Transform: A Generalized Fourier Transform",988,387000,1604793600,57,4,3,71],["jNC0jxb0OxE","Fourier Analysis: Overview",449,377000,1604793600,57,4,3,71],["YwodGM2eoy4","Model Predictive Control",733,374000,1541635200,57,4,3,71],["fy0HRViXZnE","Coding Adventure: Rubik's Cube",1462,357000,1778198400,60,5,3,71],["rRnOtKlg4jA","Coding Adventure: Additive Synthesis",2130,365000,1775606400,60,4,3,71],["HGkuRp5HfH8","Simulating 256 Bytes of RAM",1549,493000,1762560000,60,5,3,71],["iA6wRgwl7k0","Coding Adventure: Sound (and the Fourier Transform)",2611,717000,1731024000,60,4,3,71],["M3iI2l0ltbE","Coding Adventure: Marching Cubes",378,1100000,1573171200,60,5,3,71],["Cp5WWtMoeKg","Coding Adventure: Ray Marching",306,1200000,1573171200,60,5,3,71],["ahnGLM-RC1Y","A Survey of Techniques for Maximizing LLM Performance",2732,224000,1731024000,77,2,3,71],["zxQyTK8quyY","Transformer Neural Networks, ChatGPT's foundation, Clearly Explained!!!",2175,1200000,1699401600,3,0,3,70],["_L39rN6gz7Y","Decision and Classification Trees, Clearly Explained!!!",1088,1300000,1636329600,3,0,3,70],["CqOfi41LfDw","The Essential Main Ideas of Neural Networks",1134,1400000,1604793600,3,0,3,70],["O2L2Uv9pdDA","Naive Bayes, Clearly Explained!!!",912,1400000,1604793600,3,0,3,70],["Q81RR3yKn30","Regularization Part 1: Ridge (L2) Regression",1227,1400000,1541635200,3,0,3,70],["t5y-kS9nNxU","Build A Reasoning Model From Scratch 4: Inference Scaling 1 (Temperature, Top-p, Self-Consistency)",5832,14000,1791417600,5,0,3,70],["CepbWmGie0E","A Visual Tour of Modern LLM Architectures",2318,19000,1775606400,5,0,3,70],["lONyteDR4XE","LLM Building Blocks & Transformer Alternatives",1629,20000,1762560000,5,0,3,70],["hBBOjCiFcuo","Lesson 3: Practical Deep Learning for Coders 2022",5425,166000,1667865600,6,0,3,70],["MpZxV6DVsmM","Lesson 3: Deep Learning 2019 - Data blocks; Multi-label classification; Segmentation",7494,156000,1573171200,6,0,3,70],["JNxcznsrRb8","Lesson 2: Deep Learning 2018",7650,160000,1541635200,6,0,3,70],["FF7DRogZZdY","Stanford EE274: Data Compression I 2023 I Lecture 1 - Course Intro, Lossless Data Compression Basics",3760,22000,1731024000,7,5,3,70],["1GbDTTK3aR4","Stanford CS25: V3 I How I Learned to Stop Worrying and Love the Transformer",4838,22000,1731024000,7,0,3,70],["5tvmMX8r_OM","MIT 6.S191 (2021): Introduction to Deep Learning",3396,567000,1636329600,8,0,3,70],["JN6H4rQvwgY","MIT 6.S191 (2018): Introduction to Deep Learning",2560,531000,1541635200,8,0,3,70],["ec9IQMiJBhs","Transformers explained | The architecture behind LLMs",1188,42000,1731024000,10,0,3,70],["33Oct2hqGnk","Agents That Write Their Own Tools at Runtime — Sandhya Subramani, AWS",1240,16000,1791417600,18,2,3,70],["c1hGBoWw20A","Weight Folding, CUDA Streams, and the Bug That Made My Model Speak Backwards — Filip Makraduli",1038,16000,1791417600,18,6,3,70],["I8fs4omN1no","Hard Won Lessons from Building Effective AI Coding Agents – Nik Pash, Cline",858,19000,1767830400,18,2,3,70],["y-UGrYbJsJk","What every AI engineer needs to know about GPUs — Charles Frye, Modal",1192,23000,1762560000,18,6,3,70],["JfaLQqfXqPA","RFT, DPO, SFT: Fine-tuning with OpenAI — Ilan Bigio, OpenAI",6375,21000,1762560000,18,0,3,70],["uLrOI65XbDw","Everything you need to know about Fine-tuning and Merging LLMs: Maxime Labonne",1072,34000,1731024000,18,0,3,70],["nJphsM4obOk","Neural ODEs (NODEs) [Physics Informed Machine Learning]",1477,174000,1731024000,57,0,3,70],["xy3QyyhiuY4","Singular Value Decomposition (SVD): Matrix Approximation",894,303000,1604793600,57,4,3,70],["dNRDvLACg5Q","Turing Machines Explained - Computerphile",325,1200000,1415404800,59,5,3,70],["08mmKNLQVHU","Coding Adventure: Analyzing Audio",1288,361000,1767830400,60,4,3,70],["PIixc_svOwc","FreeCAD 26.3 Sketch Internals, All You Need To Know",240,36000,1793491200,73,7,3,70],["3vGaiArjKko","FreeCAD v1.1 Assembly: How To Create Exploded View & Import Into Techdraw Beginner Friendly Tutorial",919,48000,1770508800,73,7,3,70],["PcOo3rOyruI","Create Custom External Threads in FreeCAD with Additive Helix | Basic Beginners Lesson 47 | Tutorial",1036,52000,1762560000,73,7,3,70],["02Gv4gN117M","FreeCAD Learn the Basics of Tech Draw Workbench in 15 minutes. Technical Drawing Beginners Tutorial",923,121000,1667865600,73,7,3,70],["Br4l9YjCyRU","The Erdős Breakthrough",158,111000,1780876800,77,4,3,70],["KHZVXao4qXs","RL Course by David Silver - Lecture 7: Policy Gradient Methods",5638,319000,1478563200,78,0,3,70],["UoPei5o4fps","RL Course by David Silver - Lecture 6: Value Function Approximation",5805,304000,1446940800,78,0,3,70],["0g4j2k_Ggc4","RL Course by David Silver - Lecture 5: Model Free Control",5791,319000,1446940800,78,0,3,70],["YAlJCEDH2uY","The Central Limit Theorem, Clearly Explained!!!",455,1100000,1541635200,3,4,3,69],["vN5cNN2-HWE","Logistic Regression Details Pt1: Coefficients",1142,1100000,1541635200,3,0,3,69],["VxfUrUrrczs","Build A Reasoning Model From Scratch 7: Reinforcement Learning 2 (Tracking, KL Terms, And More)",6368,13000,1791676800,5,0,3,69],["gs-IDg-FoIQ","Finetuning Open-Source LLMs",1205,42000,1699401600,5,1,3,69],["nOxKexn3iBo","Getting Started With CUDA for Python Programmers",4676,83000,1731024000,6,6,3,69],["toUgBQv1BT8","Lesson 4: Practical Deep Learning for Coders 2022",5677,130000,1667865600,6,0,3,69],["BvHmRx14HQ8","Lesson 2 - Deep Learning for Coders (2020)",5465,140000,1604793600,6,0,3,69],["jm2hyJLFfN8","Stanford CS25: V4 I Behind the Scenes of LLM Pre-training: StarCoder Use Case",3697,18000,1731024000,7,0,3,69],["vKikt2d9PE0","Stanford Webinar with Dan Boneh - Hacking AI: Security & Privacy of Machine Learning Models",3496,30000,1636329600,7,3,3,69],["BxQw4CdxLr8","Stanford Lecture: Don Knuth—\"A Conjecture That Had To Be True\" (2017)",4022,30000,1541635200,7,4,3,69],["aP0Y1uAA-2Y","Stanford Seminar: Beyond Floating Point: Next Generation Computer Arithmetic",5488,32000,1510099200,7,5,3,69],["mU3g2YPKlsA","On the Biology of a Large Language Model (Part 1)",3245,59000,1762560000,11,3,3,69],["ut5kp56wW_4","Tree of Thoughts: Deliberate Problem Solving with Large Language Models (Full Paper Review)",1769,128000,1699401600,11,3,3,69],["_Ygf0GnlwmY","Mechanistic Interpretability - NEEL NANDA (DeepMind)",14264,144000,1699401600,13,3,3,69],["SWIKyLSUBIc","Information Theory for Language Models: Jack Morris",4693,11000,1762560000,17,0,3,69],["O1AR4iL30mg","The Magic of LLM Distillation — Rishabh Agarwal, Google DeepMind",2802,12000,1762560000,17,0,3,69],["y2W4FNAuPEA","Deep dive on LLM Inference at Scale — Harshul Jain, Audible & Tanmay Sah, Independent AI Researcher",5292,13000,1791417600,18,6,3,69],["FWMJQDH3iK0","Local Models: Trust, Control, Optimization — Carter Abdallah, NVIDIA",2601,13000,1786147200,18,1,3,69],["V-EDrhIhHzQ","Modern Post-Training: A Deep Dive  — Will Brown, Prime Intellect",2812,14000,1786147200,18,0,3,69],["WE_Gnowy3uw","Replacing 12K LoC with a 200 LoC Skill — David Gomes, Cursor",1162,15000,1778198400,18,2,3,69],["a2muGkT4WD4","Running LLMs on your iPhone: 40 tok/s Gemma 4 with MLX — Adrien Grondin, Locally AI",651,15000,1778198400,18,1,3,69],["8txf05vVVl4","Code Mode: Let the Code do the Talking - Sunil Pai, Cloudflare",1180,15000,1778198400,18,2,3,69],["qdmxApz3EJI","On Engineering AI Systems that Endure The Bitter Lesson - Omar Khattab, DSPy & Databricks",1152,19000,1762560000,18,2,3,69],["72XxWkd8Jrk","Effective agent design patterns in production — Laurie Voss, LlamaIndex",938,19000,1762560000,18,2,3,69],["Cl2B_hmg4gA","Lecture 75 [ScaleML Series] GPU Programming Fundamentals + ThunderKittens",9562,10000,1762560000,55,6,3,69],["tIPDwUepXcA","Lecture 58: Disaggregated LLM Inference",4519,8800,1762560000,55,6,3,69],["we3i5VuoPWk","Lecture 37: Introduction to SASS & GPU Microarchitecture",6640,9100,1762560000,55,6,3,69],["XQylGyG7yp8","Lecture 35: SGLang",2719,8900,1762560000,55,6,3,69],["9wNAgpX6z_4","Lecture 22: Hacker's Guide to Speculative Decoding in VLLM",4165,14000,1731024000,55,6,3,69],["_mv0q7-WF4E","Complex Analysis L01: Overview & Motivation, Complex Arithmetic, Euler's Formula & Polar Coordinates",1763,236000,1699401600,57,4,3,69],["Jt5R-Tm8cV8","Vector Calculus and Partial Differential Equations: Big Picture Overview",953,281000,1667865600,57,4,3,69],["MB6XGQWLV04","Fourier Series: Part 1",736,242000,1604793600,57,4,3,69],["s_9InuQAx-g","The Kalman Filter [Control Bootcamp]",371,243000,1510099200,57,4,3,69],["1CIpzeNxIhU","How AI Image Generators Work (Stable Diffusion / Dall-E) - Computerphile",1070,1000000,1667865600,59,0,3,69],["eis11j_iGMs","Lambda Calculus - Computerphile",760,1000000,1510099200,59,5,3,69],["py5byOOHZM8","CNN: Convolutional Neural Networks Explained - Computerphile",857,903000,1478563200,59,0,3,69],["macM_MtS_w4","Turing & The Halting Problem - Computerphile",374,919000,1415404800,59,5,3,69],["GSIDS_lvRv4","Public Key Cryptography - Computerphile",380,1000000,1415404800,59,5,3,69],["wA1KVZ1eOuA","Coding Adventure: Ray-Tracing Glass",2478,329000,1762560000,60,5,3,69],["_3cNcmli6xQ","Experimenting with Buses and Three-State Logic",1123,727000,1699401600,60,5,3,69],["9RHGLZLUuwc","Coding Adventure: Compute Shaders",325,863000,1573171200,60,6,3,69],["CQywdSdi5iA","The Model Context Protocol (MCP)",1175,257000,1762560000,75,2,3,69],["Bj9BD2D3DzA","Tracing the thoughts of a large language model",176,277000,1762560000,75,3,3,69],["78mEYaztGaw","Gemini 2.0 and the evolution of agentic AI | Oriol Vinyals",3117,99000,1762560000,78,2,3,69],["iOh7QUZGyiU","Deep Learning 1: Introduction to Machine Learning Based AI",6188,258000,1573171200,78,0,3,69],["YCzL96nL7j0","Long Short-Term Memory (LSTM), Clearly Explained",1245,1000000,1667865600,3,0,3,68],["AsNTP8Kwu80","Recurrent Neural Networks (RNNs), Clearly Explained!!!",997,1000000,1667865600,3,0,3,68],["YtebGVx-Fxw","Entropy (for data science) Clearly Explained!!!",995,922000,1636329600,3,0,3,68],["OtD8wVaFm6E","XGBoost Part 1 (of 4): Regression",1546,933000,1604793600,3,0,3,68],["3CC4N4z3GJc","Gradient Boost Part 1 (of 4): Regression Main Ideas",952,1000000,1573171200,3,0,3,68],["LsK-xG1cLYA","AdaBoost, Clearly Explained",1254,970000,1573171200,3,0,3,68],["Kdsp6soqA7o","Machine Learning Fundamentals: The Confusion Matrix",433,1000000,1541635200,3,0,3,68],["azXCzI57Yfc","StatQuest: Linear Discriminant Analysis (LDA) clearly explained.",912,931000,1478563200,3,0,3,68],["_UVHneBUBW0","Principal Component Analysis (PCA) clearly explained (2015)",1216,1000000,1446940800,3,0,3,68],["TVMyOJ_3Gxo","Build A Reasoning Model From Scratch 5: Inference Scaling 2 (Logprob Scoring, Self-Refinement)",4120,10000,1791417600,5,0,3,68],["qqt3aMPB81c","Lesson 4: Deep Learning 2019 - NLP; Tabular data; Collaborative filtering; Embeddings",6219,117000,1573171200,6,0,3,68],["cYfKQ6YG9Qo","Stanford CS25: V4 I From Large Language Models to Large Multimodal Models",4804,15000,1731024000,7,0,3,68],["zL9B3eXq0gY","Stanford CS25: V4 I Transformers that Transform Well Enough to Support Near-Shallow Architectures",4796,15000,1731024000,7,0,3,68],["-at-usqAIMc","Stanford AA228/CS238 Decision Making Under Uncertainty I Policy Gradient Estimation and Optimization",4860,16000,1731024000,7,0,3,68],["SEnXr6v2ifU","MIT 6.S191 (2020): Recurrent Neural Networks",2728,401000,1604793600,8,0,3,68],["dh8Rxhf7cLU","OpenAI’s CLIP explained! | Examples, links to code and pretrained model",888,48000,1636329600,10,0,3,68],["loaTGpqfctI","Byte Latent Transformer: Patches Scale Better Than Tokens (Paper Explained)",2175,48000,1762560000,11,3,3,68],["2zW33LfffPc","GPT-4 is here! What we know so far (Full Analysis)",2050,108000,1699401600,11,8,3,68],["-9evrZnBorM","BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",2413,113000,1573171200,11,0,3,68],["e3gBwLWAerw","He won a Nobel here for AlphaFold. Then he left. - John Jumper",3186,40000,1783468800,13,3,3,68],["pE3KKUKXcTM","@Asianometry  & Dylan Patel — How the semiconductor industry actually works",7854,377000,1731024000,14,6,3,68],["sYgE4ppDFOQ","Code World Model: Building World Models for Computation – Jacob Kahn, FAIR Meta",1001,15000,1767830400,18,3,3,68],["jQcsVk0KWiQ","A Taxonomy for Next-gen Reasoning — Nathan Lambert, Allen Institute (AI2) & Interconnects.ai",1161,16000,1762560000,18,3,3,68],["wsFd22SL1s8","Arrakis: How To Build An AI Sandbox From Scratch - Abhishek Bhardwaj, OpenAI",2418,17000,1762560000,18,2,3,68],["eJOjdjO45Sc","How Deep Research Works - Mukund Sridhar & Aarush Selvan, Google DeepMind",915,17000,1762560000,18,2,3,68],["Jew4lhAiqnw","Mega Lecture 91: Reinforcement Learning, Agents & OpenEnv",12111,6900,1770508800,55,2,3,68],["TORt20_HjMY","Gauss's Divergence Theorem",1613,231000,1667865600,57,4,3,68],["toj_IoCQE-4","The Fast Fourier Transform Algorithm",618,208000,1604793600,57,5,3,68],["nyqJJdhReiA","Linear Systems [Control Bootcamp]",1487,218000,1510099200,57,4,3,68],["XETZoRYdtkw","How Computer Memory Works - Computerphile",856,864000,1415404800,59,5,3,68],["Mv9NEXX1VHc","What on Earth is Recursion? - Computerphile",580,781000,1415404800,59,5,3,68],["WP-Bm65Q-1Y","Procedural Landmass Generation (E02: Noise Map)",914,687000,1478563200,60,5,3,68],["nhiFx28e7JY","A* Pathfinding (E02: node grid)",1402,678000,1446940800,60,5,3,68],["KzLWKKBtnxU","FreeCAD 0.20 For Beginners | 24 | Lid & Box / Enclosure | Threaded Holes Cross File Sub Shape Binder",2244,91000,1699401600,73,7,3,68],["ZgGIULSFlw0","FreeCAD 0.20 For Beginners | 5.1 | Tracing from Photos Project | Repeating Features | Polar Patterns",1488,99000,1667865600,73,7,3,68],["7ArmBVF2dCs","Linear Regression, Clearly Explained!!!",1647,760000,1699401600,3,4,3,67],["IN2XmBhILt4","Neural Networks Pt. 2: Backpropagation Main Ideas",1054,823000,1604793600,3,0,3,67],["0oc49DyA3hU","Hypothesis Testing and The Null Hypothesis, Clearly Explained!!!",881,887000,1604793600,3,4,3,67],["g9c66TUylZ4","Regression Trees, Clearly Explained!!!",1353,851000,1573171200,3,0,3,67],["HVXime0nQeI","StatQuest: K-nearest neighbors, Clearly Explained",330,827000,1510099200,3,0,3,67],["PaFPbb66DxQ","The Main Ideas of Fitting a Line to Data (The Main Ideas of Least Squares and Linear Regression.)",562,862000,1510099200,3,4,3,67],["A1eUVxscNq8","L19.4.3 Multi-Head Attention",457,33000,1636329600,5,0,3,67],["_rXzeWq4C6w","Lesson 5: Practical Deep Learning for Coders 2022",6169,100000,1667865600,6,0,3,67],["5L3Ao5KuCC4","Lesson 3 - Deep Learning for Coders (2020)",7583,103000,1604793600,6,0,3,67],["4u8FxNEDUeg","Lesson 8 (2019) - Deep Learning from the Foundations",7601,95000,1573171200,6,0,3,67],["blyXCk4sgEg","Intro to Machine Learning: Lesson 2",5691,97000,1541635200,6,0,3,67],["Z0ssNAbe81M","Lesson 8: Deep Learning Part 2 2018 - Single object detection",7275,94000,1541635200,6,0,3,67],["gbceqO8PpBg","Lesson 4: Deep Learning 2018",8072,95000,1541635200,6,0,3,67],["9C06ZPF8Uuc","Lesson 3: Deep Learning 2018",8198,104000,1541635200,6,0,3,67],["e3aM6XTekJc","Lesson 2: Practical Deep Learning for Coders",6857,97000,1510099200,6,0,3,67],["pC4zRb_5noQ","Stanford CS25: V1 I Transformer Circuits, Induction Heads, In-Context Learning",3574,22000,1667865600,7,3,3,67],["nR74lBO5M3s","Stanford Seminar: Google's Multilingual Neural Machine Translation System",4764,23000,1510099200,7,0,3,67],["d02VkQ9MP44","MIT 6.S191: Recurrent Neural Networks, Transformers, and Attention",3454,115000,1775606400,8,0,3,67],["to-lHJfK4pw","MIT 6.S191 (2025): Reinforcement Learning",3720,129000,1762560000,8,0,3,67],["3G5hWM6jqPk","MIT 6.S191 (2023): Deep Generative Modeling",3592,323000,1699401600,8,0,3,67],["iaSUYvmCekI","MIT 6.S191 (2020): Convolutional Neural Networks",2240,318000,1604793600,8,0,3,67],["796GVFTFiB0","Reducing Hallucinations: Merlin-Arthur Training and Evaluation EXPLAINED",1977,12000,1791417600,10,3,3,67],["jE9jAZC42NE","Were RNNs All We Needed? (Paper Explained)",1668,60000,1731024000,11,3,3,67],["r_UBBfTPcF0","Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attention",2237,60000,1731024000,11,3,3,67],["7UkJPwz_N_0","V-JEPA: Revisiting Feature Prediction for Learning Visual Representations from Video (Explained)",3003,64000,1731024000,11,3,3,67],["mwO6v4BlgZQ","Mixtral of Experts (Paper Explained)",2072,67000,1731024000,11,3,3,67],["ec56a8wmfRk","Retentive Network: A Successor to Transformer for Large Language Models (Paper Explained)",1706,103000,1699401600,11,3,3,67],["E5OnoYF2oAk","LLaMA: Open and Efficient Foundation Language Models (Paper Explained)",2467,93000,1699401600,11,3,3,67],["jSdHmImyUjk","JEPA - A Path Towards Autonomous Machine Intelligence (Paper Explained)",3578,112000,1667865600,11,3,3,67],["QWidx8cYVRs","The Mathematical Foundations of Intelligence [Professor Yi Ma]",3928,37000,1767830400,13,4,3,67],["14DXtvRJeNU","How Do AI Models Actually Think? [Dr. Laura Ruis]",4682,40000,1762560000,13,3,3,67],["X_ZVSPcZhtw","What rebuilding AlphaGo teaches us about self-play, RL, and future of LLMs - Eric Jang",9438,180000,1780876800,14,0,3,67],["64lXQP6cs5M","Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken",8642,214000,1762560000,14,3,3,67],["NdrSPm6NCdk","Evals in AI: A Deep Dive — Tejas Kumar, IBM",3589,9200,1791417600,18,2,3,67],["9JFGohx4E7U","GLM-5.2: Open Weights, Near-Frontier Intelligence — Zixuan Li, Z.ai",827,9400,1791417600,18,8,3,67],["btxG75rNJC4","Two Bugs That Hid in Plain Sight: A vLLM Debugging Detective Story — Asaf Gardin & Yuval Belfer",1086,9000,1791417600,18,6,3,67],["lRa9sPaMyy4","Your Code Has Bugs. Lean4 Has Proofs: Formal Verification for Engineers — Varun Pant, AWS",607,9400,1788825600,18,5,3,67],["ITbB9nPCX04","Run A Local LLM Across Multiple Computers! (vLLM Distributed Inference)",1005,37000,1762560000,28,6,3,67],["gzZihJ5miZE","DeepSeekR1 - Full Breakdown",1369,45000,1762560000,29,0,3,67],["WRAQYXBA_Qc","Lecture 116: GPU Kernel Formal Verification",1835,4800,1791417600,55,5,3,67],["MVh_guNbWMA","Lecture 103: Fundamentals of CuTe Layout Algebra and Category-theoretic Interpretation",9237,5300,1778198400,55,6,3,67],["zEuwuCTEf_0","Lecture 12: Flash Attention",4334,9800,1731024000,55,6,3,67],["SGhfUhlowB4","Lecture 8: CUDA Performance Checklist",4090,10000,1731024000,55,6,3,67],["5tmGKTNW8DQ","AI’s Hardware Problem",1007,673000,1699401600,56,6,3,67],["5Ge2RcvDlgw","The Extreme Engineering of ASML’s EUV Light Source",1040,698000,1667865600,56,6,3,67],["ihz2WY-E2C8","Designing Billions of Circuits with Code",731,702000,1636329600,56,6,3,67],["LqNqqidw2mg","Stokes' Theorem and Green's Theorem",1434,186000,1667865600,57,4,3,67],["yXD5IlDstNk","The Gradient Operator in Vector Calculus: Directions of Fastest Change & the Directional Derivative",904,173000,1667865600,57,4,3,67],["0iqz4tcKN58","Q-Learning: Model Free Reinforcement Learning and Temporal Difference Learning",2135,178000,1667865600,57,0,3,67],["i7q8bISGwMQ","Reinforcement Learning Series: Overview of Methods",1297,176000,1667865600,57,0,3,67],["FcXZ28BX-xE","Shannon Nyquist Sampling Theorem",1039,177000,1636329600,57,4,3,67],["y7KLbd7n75g","Wavelets and Multiresolution Analysis",912,187000,1604793600,57,4,3,67],["jVYs-GTqm5U","The Fourier Transform",876,186000,1604793600,57,4,3,67],["u5Sv7YKAkt4","Controllability [Control Bootcamp]",1950,186000,1510099200,57,4,3,67],["h7nJ6ZL4Lf0","Stability and Eigenvalues [Control Bootcamp]",1170,185000,1510099200,57,4,3,67],["wL22URoMZjo","Sleeper Agents in Large Language Models - Computerphile",818,292000,1762560000,59,3,3,67],["86cQJ0MMses","TLS Handshake Explained - Computerphile",1019,661000,1604793600,59,5,3,67],["Ml3-kVYLNr8","Von Neumann Architecture - Computerphile",980,729000,1541635200,59,5,3,67],["NF1pwjL9-DE","Elliptic Curves - Computerphile",522,653000,1541635200,59,4,3,67],["t1e8gqXLbsU","What is a Monad? - Computerphile",1310,665000,1541635200,59,5,3,67],["Sw9r8CL98N0","Generative Adversarial Networks (GANs) - Computerphile",1281,682000,1510099200,59,0,3,67],["thrx3SBEpL8","Why Use Binary? - Computerphile",509,685000,1478563200,59,5,3,67],["Q2aEzeMDHMA","JPEG DCT, Discrete Cosine Transform (JPEG Pt2)- Computerphile",912,698000,1446940800,59,5,3,67],["Eysf6-E3ino","Coding Adventure: Improving my Rubik's Cube Solver",1648,191000,1780876800,60,5,3,67],["YCdvL87bv0w","Things you should know about ABS",648,548000,1541635200,69,7,3,67],["tw4iP0luFVU","Designing Buckles, Clips and Snaps for 3D Printing - Detailed Guide",2000,445000,1573171200,70,7,3,67],["_BuuGswqWWE","How to Set your Nozzle Height and Level your Bed - 3D Printing 101",626,419000,1478563200,70,7,3,67],["1mGI9vfsCCA","Part vs Part Design which one to choose?  Basic Beginners FreeCAD v1 / 0.22 | Lesson 5",447,48000,1731024000,73,7,3,67],["jfhgcDcAAlk","FreeCAD Basics of Using Sketch Attachment Modes",1551,69000,1699401600,73,7,3,67],["B0gXZv_3L-E","FreeCAD 0.20 For Beginners | 16 | Assembly with Fasteners | Assembly 4 WB",2022,82000,1667865600,73,7,3,67],["W-whnGUUQ2Y","FreeCAD 0.20 For Beginners | 9 | Modelling Lipped Containers | Multibody Workflow | Part Design",1693,75000,1667865600,73,7,3,67],["WZZK4vmXK9Q","FreeCAD 0.20 For Beginners | 5.2 | Tracing from Photos Project | Finishing Our Model",2126,75000,1667865600,73,7,3,67],["j2knrqAzYVY","Translating Claude’s thoughts into language",197,162000,1778198400,75,3,3,67],["ISk80iLhdfU","Reinforcement Learning 1: Introduction to Reinforcement Learning",6197,198000,1573171200,78,0,3,67],["CDnkFbW-uEQ","MoE Token Routing Explained: How Mixture of Experts Works (with Code)",2055,9500,1770508800,79,0,3,67],["EAd4oQtEJOM","Reinforcement Learning from Human Feedback  From Zero to ChatGPT [Record of the live]",3638,22000,1699401600,79,0,3,67],["4MmSZrAlqKc","Imaginary Numbers Are Real [Part 13: Riemann Surfaces]",829,1200000,1478563200,2,4,3,66],["viZrOnJclY0","Word Embedding and Word2Vec, Clearly Explained!!!",972,620000,1699401600,3,0,3,66],["Xz0x-8-cgaQ","Bootstrapping Main Ideas!!!",567,641000,1636329600,3,4,3,66],["qtaqvPAeEJY","Covariance, Clearly Explained!!!",1343,674000,1573171200,3,4,3,66],["SzZ6GpcfoQY","Calculating the Mean, Variance and Standard Deviation, Clearly Explained!!!",862,643000,1573171200,3,4,3,66],["NGf0voTMlcs","Regularization Part 2: Lasso (L1) Regression",499,734000,1541635200,3,0,3,66],["zITIFTsivN8","Multiple Regression, Clearly Explained!!!",325,672000,1510099200,3,0,3,66],["oI3hZJqXJuc","The Main Ideas behind Probability Distributions",315,661000,1510099200,3,4,3,66],["0PjHri8tc1c","L19.4.2 Self-Attention and Scaled Dot-Product Attention",969,25000,1636329600,5,0,3,66],["p50s63nPq9I","Lesson 4 - Deep Learning for Coders (2020)",7295,78000,1604793600,6,0,3,66],["9spwoDYwW_I","Lesson 7: Deep Learning 2019 - Resnets from scratch; U-net; Generative (adversarial) networks",7926,80000,1573171200,6,0,3,66],["CJKnDu2dxOE","Lesson 5: Deep Learning 2019 - Back propagation; Accelerated SGD; Neural net from scratch",8014,88000,1573171200,6,0,3,66],["RcJ1YXHLv5o","Stanford CS25: V4 I Demystifying Mixtral of Experts",3872,11000,1731024000,7,0,3,66],["9Th86_3Ea7o","Stanford Seminar - ML Explainability Part 2 I Inherently Interpretable Models",2698,20000,1667865600,7,3,3,66],["QNnayf--_yk","Statistical Learning: 8.1 Tree based methods",878,19000,1667865600,7,0,3,66],["lLlG5xkyqIA","Statistical Learning: 6.6 Shrinkage methods and ridge regression",758,18000,1667865600,7,4,3,66],["SdTZAMDKrNY","MIT 6.S191 (2025): Deep Generative Modeling",2937,123000,1762560000,8,0,3,66],["2xqkSUhmmXU","MIT 6.S191 (2024): Convolutional Neural Networks",4078,164000,1731024000,8,0,3,66],["NmLK_WQBxB4","MIT 6.S191 (2023): Convolutional Neural Networks",3315,277000,1699401600,8,0,3,66],["QvkQ1B3FBqA","MIT 6.S191 (2022): Recurrent Neural Networks and Transformers",3498,260000,1667865600,8,0,3,66],["qjrad0V0uJE","MIT 6.S191 (2021): Recurrent Neural Networks",3631,301000,1636329600,8,0,3,66],["BprirYymXrg","LLM Lecture: A Deep Dive into Transformers, Prompts, and Human Feedback",5469,15000,1762560000,10,0,3,66],["Dp6iICL2dVI","Masked Autoencoders Are Scalable Vision Learners – Paper explained and animated!",776,33000,1667865600,10,0,3,66],["zKohTkN0Fyk","[Paper Analysis] On the Theoretical Limitations of Embedding-Based Retrieval (Warning: Rant)",2937,37000,1762560000,11,3,3,66],["hpC4qjWu_aY","Context Rot: How Increasing Input Tokens Impacts LLM Performance (Paper Analysis)",2269,33000,1762560000,11,3,3,66],["RAEy3JZmIaA","Energy-Based Transformers are Scalable Learners and Thinkers (Paper Review)",2871,33000,1762560000,11,3,3,66],["x8pW19wKfXQ","RWKV: Reinventing RNNs for the Transformer Era (Paper Explained)",3737,82000,1699401600,11,3,3,66],["VQoyypYTz2U","All about AI Accelerators: GPU, TPU, Dataflow, Near-Memory, Optical, Neuromorphic & more (w/ Author)",3755,89000,1667865600,11,6,3,66],["dND-7llwrpw","Grokking: Generalization beyond Overfitting on small algorithmic datasets (Paper Explained)",1787,86000,1636329600,11,0,3,66],["GWt6Fu05voI","[Classic] Deep Residual Learning for Image Recognition (Paper Explained)",1881,85000,1604793600,11,3,3,66],["unUeI7e-iVs","Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay",5525,6200,1770508800,17,3,3,66],["PAz_-xPJcRM","The RLVR Revolution — with Nathan Lambert (AI2, Interconnects.ai)",4740,7100,1762560000,17,0,3,66],["pOvWgX7IJsc","Can LLMs Write Fast Multi-GPU Kernels? — Simran Arora, Together AI",1800,7500,1788825600,18,6,3,66],["X4dEHRzBLmc","Judge the Judge: Building LLM Evaluators That Actually Work with GEPA — Mahmoud Mabrouk, Agenta AI",2451,8900,1778198400,18,2,3,66],["k1t2xyWMUdY","How METR measures Long Tasks and Experienced Open Source Dev Productivity - Joel Becker, METR",4552,11000,1770508800,18,3,3,66],["TZnJYRTSGVk","Lecture 113: Every Microsecond Matters: Achieving Near Speed-of-Light Latency in GPU Collectives",3173,3800,1788825600,55,6,3,66],["ErTmTCRP1_U","Lecture 45: Outperforming cuBLAS on H100",4561,5400,1762560000,55,6,3,66],["HxyM2Chu9Vc","Nvidia's Computational Lithography Breakthrough",910,355000,1731024000,56,6,3,66],["SB8qIO6Ti_M","How ASML Won Lithography (& Why Japan Lost)",954,654000,1636329600,56,6,3,66],["w1UsKanMatM","Residual Networks (ResNet) [Physics Informed Machine Learning]",1046,103000,1731024000,57,0,3,66],["sJIFUTITfBc","Model Based Reinforcement Learning: Policy Iteration, Value Iteration, and Dynamic Programming",1630,163000,1667865600,57,0,3,66],["KmQkDgu-Qp0","Deep Learning to Discover Coordinates for Dynamics: Autoencoders & Physics Informed Machine Learning",1616,157000,1636329600,57,0,3,66],["zIQpxmLwbXQ","Turbulence Closure Models: Reynolds Averaged Navier Stokes (RANS) & Large Eddy Simulations (LES)",2014,150000,1636329600,57,4,3,66],["IUiKAD6cuTA","Deep Reinforcement Learning: Neural Networks for Learning Control Laws",1275,169000,1636329600,57,0,3,66],["g-eNeXlZKAQ","Inner Products in Hilbert Space",521,163000,1604793600,57,4,3,66],["WmDnaoY2Ivs","Singular Value Decomposition (SVD): Dominant Correlations",680,170000,1604793600,57,4,3,66],["oulLR06lj_E","Data-Driven Control: Overview",1471,159000,1541635200,57,4,3,66],["PrfxmkBsYKE","Cayley-Hamilton Theorem [Control Bootcamp]",417,156000,1510099200,57,4,3,66],["k_kyepATqB8","Shor's Algorithm for Quantum Computing - Computerphile",2339,171000,1786147200,59,5,3,66],["jsraR-el8_o","Hash Collisions & The Birthday Paradox - Computerphile",861,241000,1762560000,59,4,3,66],["KcSXcpluDe4","How AI 'Understands' Images (CLIP) - Computerphile",1085,382000,1731024000,59,0,3,66],["Yjrfm_oRO0w","Diffie Hellman -the Mathematics bit- Computerphile",425,572000,1541635200,59,4,3,66],["vQ60rFwh2ig","The True Power of the Matrix (Transformations in Graphics) - Computerphile",886,584000,1415404800,59,4,3,66],["mZfyt03LDH4","A* Pathfinding (E03: algorithm implementation)",1451,473000,1446940800,60,5,3,66],["BA12Z7gQ4P0","Using an EEPROM to replace combinational logic",1522,1500000,1510099200,61,5,3,66],["er7q-CJL1lc","Klipper input shaping - A leap forward in high speed AND high quality 3D printing [Rat Rig part 4]",1277,455000,1636329600,68,7,3,66],["Mnvj6xCzikM","Remove ringing by tuning acceleration and junction deviation - step by step guide",1039,452000,1604793600,68,7,3,66],["Kf907duwAAE","Build your own 3D Printer: Which hotend to pick!",658,446000,1478563200,69,7,3,66],["8hhNiJ2_qgE","FreeCAD: How To Structure Your Projects & Share Across Multiple Files | Basic Beginners Lesson 55",1472,24000,1770508800,73,7,3,66],["i6V4xZOZjR0","FreeCAD 0.20 For Beginners | 10 | Curved surfaces using Surface Workflow | Part Design",1451,62000,1667865600,73,7,3,66],["eP_pMWu25tM","FreeCAD 0.20 For Beginners | 7 | Revolve and Sweep for More Complex Modelling | Part Design",1484,70000,1667865600,73,7,3,66],["lLCqwrhGM3o","FreeCAD 3D to 2D Export in DXF for Laser Cut or CNC. Learn the basics for Beginners",839,61000,1667865600,73,7,3,66],["Zu2oET6Xjow","The Future of Math with o1 Reasoning with Terence Tao, Mark Chen, and James Donovan",5167,69000,1762560000,77,3,3,66],["ItMutbeOHtc","RL Course by David Silver - Lecture 8: Integrating Learning and Planning",6013,153000,1446940800,78,0,3,66],["jlGf2qieSk0","RoPE: Understanding Rotary Positional Embeddings in transformers",1120,7000,1778198400,79,0,3,66],["qx7hirqgfuU","Why Deep Learning Works Unreasonably Well [How Models Learn Part 3]",2049,402000,1762560000,2,0,3,65],["z5IG_6_zPDo","Imaginary Numbers Are Real [Part 6: The Complex Plane]",218,1000000,1446940800,2,4,3,65],["RDZUdRSDOok","Clustering with DBSCAN, Clearly Explained!!!",570,547000,1667865600,3,0,3,65],["JQc3yx0-Q9E","How to calculate p-values",1515,559000,1604793600,3,4,3,65],["vMh0zPT0tLI","Stochastic Gradient Descent, Clearly Explained!!!",653,605000,1573171200,3,0,3,65],["Dn6b9fCIUpM","Maximum Likelihood For the Normal Distribution, step-by-step!!!",1190,632000,1541635200,3,4,3,65],["BfKanl1aSG0","Logistic Regression Details Pt 2: Maximum Likelihood",623,559000,1541635200,3,0,3,65],["NEaUSP4YerM","StatQuest: t-SNE, Clearly Explained",708,579000,1510099200,3,0,3,65],["7xHsRkOdVwo","StatQuest: Hierarchical Clustering",679,577000,1510099200,3,0,3,65],["tstbZXNCfLY","L19.5.1 The Transformer Architecture",1356,24000,1636329600,5,0,3,65],["6StU6UtZEbU","Lesson 10: Deep Learning Foundations to Stable Diffusion, 2022",6554,71000,1667865600,6,0,3,65],["AdhG64NF76E","Lesson 6: Practical Deep Learning for Coders 2022",6175,66000,1667865600,6,0,3,65],["hkBa9pU-H48","Lesson 6: Deep Learning 2019 - Regularization; Convolutions; Data ethics",8263,72000,1573171200,6,0,3,65],["nG3tT31nPmQ","Lesson 14: Deep Learning Part 2 2018 - Super resolution; Image segmentation with Unet",7370,75000,1541635200,6,0,3,65],["_6n8r523QP8","Stanford Seminar - ML Explainability Part 3 I Post hoc Explanation Methods",4358,17000,1667865600,7,3,3,65],["0tfPuddPhEY","Statistical Learning: 6.7 The Lasso",922,17000,1667865600,7,4,3,65],["pqIcoskUuWs","MIT 6.S191: Convolutional Neural Networks",3406,84000,1778198400,8,0,3,65],["5v1JnYv_yWs","MIT 6.S191 (2019): Introduction to Deep Learning",2728,234000,1573171200,8,0,3,65],["O_eUzrFU6eQ","Stealing Part of a Production LLM | API protects LLMs no more",1129,18000,1731024000,10,3,3,65],["xqDeAz0U-R4","Imagen, the DALL-E 2 competitor from Google Brain, explained 🧠| Diffusion models illustrated",904,30000,1667865600,10,0,3,65],["QqejV0LNDHA","ConvNeXt: A ConvNet for the 2020s – Paper Explained (with animations)",1160,30000,1667865600,10,0,3,65],["DwaBQbqh5aE","Self-Attention with Relative Position Representations – Paper explained",618,28000,1636329600,10,0,3,65],["0OaEv1a5jUM","xLSTM: Extended Long Short-Term Memory",3420,45000,1731024000,11,3,3,65],["FY5j3P9tCeA","Text Embeddings Reveal (Almost) As Much As Text",2226,42000,1731024000,11,3,3,65],["KwpeuqT69fw","Scalable Extraction of Training Data from (Production) Language Models (Paper Explained)",2858,44000,1731024000,11,3,3,65],["eyxmSmjmNS0","[Classic] Generative Adversarial Networks (Paper Explained)",2224,73000,1604793600,11,3,3,65],["c_hmxRVDXBE","The Weird ChatGPT Hack That Leaked Training Data [Dr. Yannic Kilcher / Prof. Florian Tramer]",894,31000,1762560000,13,5,3,65],["YEUclZdj_Sc","Why next-token prediction is enough for AGI - Ilya Sutskever (OpenAI Chief Scientist)",128,217000,1731024000,14,0,3,65],["wH8YKCE82Qo","⚡ Open Model Pretraining Masterclass — Elie Bakouch, HuggingFace SmolLM 3, FineWeb, FinePDF",3820,5800,1762560000,17,0,3,65],["woIYJYd_etI","What Is an Inference Engine, Anyway? — Charles Frye, Modal",3599,6900,1791417600,18,6,3,65],["bRGyYaE0lxI","GPU Died. Training Didn't: Self-Healing Training at Scale — Crusoe",1014,6600,1791417600,18,6,3,65],["ESbWpPT_9-o","Run Frontier AI at Home — Alex Cheema, EXO Labs",6302,8000,1780876800,18,1,3,65],["YZHPEkfy2kc","$1 AI Guardrails: The Unreasonable Effectiveness of Finetuned ModernBERTs – Diego Carpintero",2633,7600,1778198400,18,0,3,65],["Ahtaha9fEM0","Introduction to LLM serving with SGLang - Philip Kiely and Yineng Zhang, Baseten",2622,9100,1762560000,18,6,3,65],["0NHCyq8bBcM","Remote MCPs: What we learned from shipping — John Welsh, Anthropic",912,9000,1762560000,18,2,3,65],["E0k9Ppq6yXY","Stateful Agents — Full Workshop with Charles Packer of Letta and MemGPT",4774,10000,1762560000,18,2,3,65],["qBHfQT3YtyY","Lessons From A Year Building With LLMs",2121,14000,1731024000,18,2,3,65],["RYZ0FMAKRFs","Mistral 8x7B Part 1- So What is a Mixture of Experts Model?",753,45000,1731024000,29,0,3,65],["7XsSd9mqay4","Lecture 115: Proving Kernels Correct Instead of Testing Them",2407,3200,1791417600,55,5,3,65],["loZ4xQ5RZuU","Lecture 112: Production Megakernels for Real-World Inference",3485,3500,1788825600,55,6,3,65],["njgow_zaJMw","Lecture 29: Triton Internals",3889,7800,1731024000,55,6,3,65],["ws7angQYIxI","Lecture 13: Ring Attention",4535,7400,1731024000,55,6,3,65],["1u9xUK3G4VM","Lecture 7 Advanced Quantization",5006,7000,1731024000,55,1,3,65],["wVsR-YhaHlM","Lecture 5: Going Further with CUDA for Python Programmers",4654,7700,1731024000,55,6,3,65],["QdESABi1Avk","How the EUV Mirrors are Made",1156,213000,1762560000,56,6,3,65],["US6YO-IK64w","Why the Chips Get Hot",1084,205000,1762560000,56,6,3,65],["pAq_dilfB_0","Complex Analysis L06: Analytic Functions and Cauchy-Riemann Conditions",2629,120000,1699401600,57,4,3,65],["Vtijyyo5fKI","Matrix Systems of Differential Equations",1498,132000,1667865600,57,4,3,65],["ebfOSDj4j3I","Taylor Series and Power Series Made Easy (with Pictures): Review of Calculus",1735,128000,1667865600,57,4,3,65],["So7vlARGs68","The Divergence of a Vector Field: Sources and Sinks",1257,125000,1667865600,57,4,3,65],["-FvrON0OmYc","The Anatomy of a Dynamical System",1073,125000,1636329600,57,4,3,65],["PjeOmOz9jSY","Linear Systems of Equations, Least Squares Regression, Pseudoinverse",713,126000,1604793600,57,4,3,65],["uOfonONtIuk","How DNS Works - Computerphile",484,506000,1604793600,59,5,3,65],["CE8UhcyJS0I","Busy Beaver Turing Machines - Computerphile",1076,484000,1415404800,59,5,3,65],["1dOCHwf8zVQ","Heartbleed, Running the Code - Computerphile",642,471000,1415404800,59,5,3,65],["goOa3DGezUA","Elegant Compression in Text (The LZ 77 Method) - Computerphile",523,522000,1383868800,59,5,3,65],["kzwT3wQWAHE","Complex Behaviour from Simple Rules: 3 Simulations",652,401000,1636329600,60,5,3,65],["bVQUSndDllU","A Brief Introduction to Neural Networks (unfinished series)",338,406000,1541635200,60,0,3,65],["4qH4unVtJkE","Twos complement: Negative numbers in binary",829,1300000,1478563200,61,5,3,65],["3hi2_9-YCbM","Master infill to take your 3D prints to the next level",1068,346000,1699401600,68,7,3,65],["tKg-jIrr_JE","Beginner’s guide to building safe 18650 batteries",1178,382000,1667865600,68,7,3,65],["ND8uJWlOgIQ","Guide: Properly picking and using MOSFETs!",548,349000,1478563200,69,7,3,65],["eure74A9HNc","3D printing guides: Everything about heated beds!",618,361000,1446940800,69,7,3,65],["Mbn1ckR86Z8","3D printing guides: Calibration and why you might be doing it wrong",430,353000,1446940800,69,7,3,65],["YUPfBJz3I6Y","3D printing guides - Calibrating your extruder",487,377000,1415404800,69,7,3,65],["-KZXH5aRcBo","FreeCAD How To Generate Multiple Variants Automatically With Spreadsheet | Basic Beginners Lesson 52",1060,21000,1765152000,73,7,3,65],["h02ti0Bl6zk","OpenAI DevDay: Keynote Recap",211,88000,1731024000,77,8,3,65],["sGuiWX07sKw","RL Course by David Silver - Lecture 9: Exploration and Exploitation",5958,133000,1446940800,78,0,3,65],["kYB8IZa5AuE","Linear transformations and matrices | Chapter 3, Essence of linear algebra",659,7300000,1478563200,1,4,3,64],["k7RM-ot2NWY","Linear combinations, span, and basis vectors | Chapter 2, Essence of linear algebra",599,7600000,1478563200,1,4,3,64],["z64a7USuGX0","What the Books Get Wrong about AI [Double Descent]",2046,323000,1762560000,2,0,3,64],["UGO_Ehywuxc","The Dark Matter of AI [Mechanistic Interpretability]",1449,316000,1762560000,2,3,3,64],["OjIwCOevUew","The Most Useful Curve in Mathematics [Logarithms]",1423,484000,1731024000,2,4,3,64],["YHvR8siIiD0","Imaginary Numbers Are Real [Part 7: Complex Multiplication]",265,901000,1446940800,2,4,3,64],["bxe2T-V8XRs","Neural Networks Demystified [Part 1: Data and Architecture]",187,838000,1415404800,2,0,3,64],["PSs6nxngL6k","Attention for Neural Networks, Clearly Explained!!!",951,480000,1699401600,3,0,3,64],["9wCnvr7Xw4E","Bayes' Theorem, Clearly Explained!!!!",840,526000,1636329600,3,4,3,64],["vP06aMoz4v8","Machine Learning Fundamentals: Sensitivity and Specificity",707,511000,1604793600,3,0,3,64],["xZ_z8KWkhXE","Pearson's Correlation, Clearly Explained!!!",1153,504000,1573171200,3,4,3,64],["vikkiwjQqfU","Population and Estimated Parameters, Clearly Explained!!!",871,519000,1573171200,3,4,3,64],["ARfXDSkQf1Y","Odds and Log(Odds), Clearly Explained!!!",691,465000,1541635200,3,4,3,64],["NF5_btOaCig","Using Linear Models for t-tests and ANOVA, Clearly Explained!!!",698,502000,1510099200,3,4,3,64],["TqOeMYtOc1w","Confidence Intervals, Clearly Explained!!!",402,467000,1446940800,3,4,3,64],["rgmJep4Sba4","Insights from Finetuning LLMs with Low-Rank Adaptation",829,12000,1731024000,5,0,3,64],["p4ZZq0736Po","Lesson 7: Practical Deep Learning for Coders 2022",6403,57000,1667865600,6,0,3,64],["cX30jxMNBUw","Lesson 6 - Deep Learning for Coders (2020)",7367,62000,1604793600,6,0,3,64],["YSFG_W8JxBo","Intro to Machine Learning: Lesson 3",5023,64000,1541635200,6,0,3,64],["0frKXR-2PBY","Lesson 9: Deep Learning Part 2 2018 - Multi-object detection",7534,59000,1541635200,6,0,3,64],["J99NV9Cr75I","Lesson 5: Deep Learning 2018",8319,63000,1541635200,6,0,3,64],["V2h3IOBDvrA","Lesson 4: Practical Deep Learning for Coders",7427,59000,1510099200,6,0,3,64],["jJb2qytbcNg","Statistical Learning: 10.1 Introduction to Neural Networks",931,13000,1667865600,7,0,3,64],["jgoa28FR__Y","Statistical Learning: 5.3 Cross Validation the wrong and right way",608,13000,1667865600,7,0,3,64],["8JVRbHAVCws","MIT 6.S191 (2024): Reinforcement Learning",3619,128000,1731024000,8,0,3,64],["taCVT5vDAk0","TiDAR: Think in Diffusion, Talk in Autoregression (Paper Analysis)",2822,24000,1767830400,11,3,3,64],["v67plFw1nMw","Titans: Learning to Memorize at Test Time (Paper Analysis)",1951,25000,1767830400,11,3,3,64],["PW4JiJ-WaY4","Beyond A*: Better Planning with Transformers via Search Dynamics Bootstrapping (Searchformer)",2645,37000,1731024000,11,3,3,64],["ZNK4nfgNQpM","AlphaGeometry: Solving olympiad geometry without human demonstrations (Paper Explained)",2127,40000,1731024000,11,3,3,64],["4Cclp6yPDuw","Scaling Transformer to 1M tokens and beyond with RMT (Paper Explained)",1474,59000,1699401600,11,3,3,64],["jltgNGt8Lpg","Neural Ordinary Differential Equations",1339,68000,1573171200,11,0,3,64],["zSAGzfspuDE","The AI Progress Chart Everyone Is Misreading — Beth Barnes & David Rein",6807,22000,1778198400,13,3,3,64],["8u2pW2zZLCs","LSTM: The Comeback Story? [Prof. Sepp Hochreiter]",4022,28000,1762560000,13,0,3,64],["Wo95ob_s_NI","John Schulman (OpenAI Cofounder) — Reasoning, RLHF, & plan for 2027 AGI",5751,193000,1731024000,14,3,3,64],["UTuuTTnjxMQ","Sholto Douglas & Trenton Bricken — How LLMs actually think",11593,199000,1731024000,14,0,3,64],["-cSSYnko63E","⚡️GPT5-Codex-Max: Training Agents with Personality, Tools & Trust — Brian Fioca + Bill Chen, OpenAI",1666,4300,1767830400,17,8,3,64],["3vFu79ccDcc","⚡️ARC-AGI-3: The Interactive Reasoning Benchmark",2381,4800,1762560000,17,3,3,64],["KjH7Gl0_pq0","DeepSeek V3, SGLang, and the state of Open Model Inference in 2025 (Quantization, MoEs, Pricing)",3438,4800,1762560000,17,1,3,64],["htod7Nv1cBc","How MiniMax M3 Was Built: Sparse Attention and Native Multimodality — Olive Song",903,6000,1792281600,18,0,3,64],["AQv3qRCG6Gw","Reinforcement Learning without Verifiable Rewards — Will Brown, Prime Intellect",1167,6200,1786147200,18,0,3,64],["VhfAVA3BG2I","How Transformers Finally Ate Vision – Isaac Robinson, Roboflow",1025,6000,1780876800,18,0,3,64],["ZuiJjkbX0Og","How LLMs work for Web Devs: GPT in 600 lines of Vanilla JS - Ishan Anand",6094,8700,1762560000,18,0,3,64],["b0xlsQ_6wUQ","The Future of Qwen: A Generalist Agent Model — Junyang Lin, Alibaba Qwen",1514,8800,1762560000,18,8,3,64],["zuMw0pkPXpU","Tool Calling Is Not Just Plumbing for AI Agents — Roy Derks",1518,8400,1762560000,18,2,3,64],["nxHKBq5ZU9U","I've Used Claude Code for 2,000+ Hours - Here's How I Build Anything With It",1368,32000,1775606400,20,2,3,64],["q_cDvCq1pww","Ollama AI Home Server ULTIMATE Setup Guide",1566,56000,1731024000,25,1,3,64],["5NJ6V8i1Xd8","Build your OWN WireGuard VPN! Here's how",741,603000,1699401600,27,5,3,64],["s1uFVfuT2aw","Train an LLM From Scratch On NVIDIA Jetson Nano (Step-by-Step Guide)",2867,23000,1762560000,28,0,3,64],["zSn8il5Mo5s","The Rise of Oracle, SQL and the Relational Database",1339,278000,1731024000,56,5,3,64],["z8L202FlmD4","The Birth of SQL & the Relational Database",1208,258000,1731024000,56,5,3,64],["m-8G1Yixb34","The History of the FPGA: The Ultimate Flex",1088,438000,1667865600,56,5,3,64],["TRZqE6H-dww","How Nvidia Won Graphics Cards",1362,437000,1636329600,56,6,3,64],["jJIO7aRXUCg","How ASML Builds a $150 Million EUV Machine",892,464000,1636329600,56,6,3,64],["W8PybqAk6Ik","Fourier Neural Operator (FNO) [Physics Informed Machine Learning]",1066,68000,1731024000,57,0,3,64],["fiX8c-4K0-Q","AI/ML+Physics Part 3: Designing an Architecture [Physics Informed Machine Learning]",2190,68000,1731024000,57,0,3,64],["Vx2DpMgplEM","A Machine Learning Primer: How to Build an ML Model",828,67000,1731024000,57,0,3,64],["CpKD9aYEdms","Complex Analysis L04: The Complex Logarithm, Log(z)",1681,97000,1699401600,57,4,3,64],["HOWJp8NV5xU","Runge-Kutta Integrator Overview:  All Purpose Numerical Integration of Differential Equations",1822,113000,1699401600,57,4,3,64],["RCWkzzLgwf0","Linearizing Nonlinear Differential Equations Near a Fixed Point",1408,116000,1667865600,57,4,3,64],["VjWtMl6vQ3Q","PDE 101: Separation of Variables! ...or how I learned to stop worrying and solve Laplace's equation",2994,118000,1667865600,57,4,3,64],["nmvs0vrBT18","Laplace's Equation and Poisson's Equation",1075,114000,1667865600,57,4,3,64],["QtiCZQIwBT8","The Curl of a Vector Field: Measuring Rotation",1609,122000,1667865600,57,4,3,64],["-hO-AnFYm6M","Nonlinear Control: Hamilton Jacobi Bellman (HJB) and Dynamic Programming",1059,121000,1667865600,57,4,3,64],["wDVteayWWvU","Overview of Deep Reinforcement Learning Methods",1490,113000,1667865600,57,0,3,64],["sQvrK8AGCAo","Dynamic Mode Decomposition (Overview)",1098,117000,1541635200,57,4,3,64],["Kap3TZwAsv0","Data-Driven Dynamical Systems Overview",1303,118000,1541635200,57,4,3,64],["BmBw827GgFA","Mutex Explained - Computerphile",702,120000,1793923200,59,5,3,64],["fzlflyw7X2I","The Next Big SHA? SHA3 Sponge Function Explained - Computerphile",1657,170000,1762560000,59,5,3,64],["-ShwJqAalOk","Breaking RSA - Computerphile",890,422000,1667865600,59,5,3,64],["s22eJ1eVLTU","What are Digital Signatures? - Computerphile",617,420000,1636329600,59,5,3,64],["_8yVOC4ciXc","GPT3: An Even Bigger Language Model - Computerphile",1557,445000,1604793600,59,0,3,64],["lNa9bQRPMB8","MegaProcessor - Computerphile",489,388000,1510099200,59,5,3,64],["BFdMrDOx_CM","Inside a Neural Network - Computerphile",942,439000,1478563200,59,0,3,64],["l42lr8AlrHk","Deep Learning - Computerphile",666,439000,1478563200,59,0,3,64],["JyxRH18YlpA","How Bitcoin Works - Computerphile",450,414000,1478563200,59,5,3,64],["sRFM5IEqR2w","Canny Edge Detector - Computerphile",471,395000,1478563200,59,5,3,64],["UvI-AMAtrvE","AND OR NOT - Logic Gates Explained - Computerphile",521,433000,1446940800,59,5,3,64],["lKTsv6iVxV4","Binary: Plusses & Minuses (Why We Use Two's Complement) - Computerphile",976,435000,1446940800,59,5,3,64],["M5c_RFKVkko","Entropy in Compression - Computerphile",732,407000,1383868800,59,5,3,64],["XE4VP_8Y0BU","Quick Sort - Computerphile",203,436000,1383868800,59,5,3,64],["Lto-ajuqW3w","Compression - Computerphile",458,440000,1383868800,59,5,3,64],["MRNFcywkUSA","Procedural Landmass Generation (E03: Octaves)",719,388000,1478563200,60,5,3,64],["3XDH-fZKnQk","Let's build a digital clock!",1607,326000,1788825600,61,7,3,64],["mhBya3JYteQ","How do aircraft systems communicate?",1308,438000,1762560000,61,5,3,64],["AHYNxpqKqwo","The RS-232 protocol",1570,1000000,1667865600,61,5,3,64],["g_koa00MBLg","How do hardware timers work?",1877,1000000,1667865600,61,5,3,64],["oO8_2JJV0B4","Assembly language vs. machine code — 6502 part 3",935,1100000,1604793600,61,5,3,64],["Lm_uuB9_2yg","How to modify an STL or mesh - 3D design for 3D printing pt7",830,317000,1667865600,68,7,3,64],["yP7_qZ_Mi78","Designing STRONG parts: tips and tricks - 3D design for 3D printing pt5",1064,289000,1667865600,68,7,3,64],["-Gdk0wHg51w","Bigtreetech SKR V1.4 (Turbo) comprehensive guide",1158,337000,1604793600,68,7,3,64],["O5V8iEjPiR0","3D printing engineering parts: PLA vs ASA vs PC vs PP vs nylon vs tough resin",993,289000,1573171200,68,7,3,64],["GNGN2iSQ5j4","SKR V1.3 Comprehensive guide - 32bit 3D printing for $20",1424,295000,1573171200,68,7,3,64],["7VHwcEroHPk","TMC2208 guide - Stepper driver upgrades part 2",1309,335000,1573171200,68,7,3,64],["9mohplbz7Wg","FreeCAD How To Generate Multiple Variants Without Re-Modeling | Basic Beginners Lesson 53  (FCv1.1) ",890,18000,1767830400,73,7,3,64],["KSHiOTE_Ibo","Custom Edges in FreeCAD: Fillets, Chamfers, and Pipes | Basic Beginners FreeCAD Lesson 50",761,18000,1762560000,73,7,3,64],["A9Lx-d-l42c","Total Assembly Control in FreeCAD With One Master Sketch and VarSet",1450,20000,1762560000,73,7,3,64],["87OVuYDrOIU","FreeCAD Basics: What are Local and Global Coordinates and Why Is It Important?",1019,42000,1699401600,73,7,3,64],["B31MCCWDB9k","FreeCAD 0.20 For Beginners | 8 | Part vs Part Design Revolve Workflows | When and Where to use",2087,51000,1667865600,73,7,3,64],["ij-W5ugn2uw","FreeCAD: Learn How To Build a Spreadsheet Driven Model in 15 minutes. Quick and Easy for Beginners",1097,51000,1667865600,73,7,3,64],["sAitLFLbgDA","Build Hour: AgentKit",2779,48000,1762560000,77,2,3,64],["aQJP3Z2Ho8U","DeepMind x UCL RL Lecture Series - Exploration & Control [2/13]",7814,108000,1636329600,78,0,3,64],["FBggC-XVF4M","DeepMind x UCL | Deep Learning Lectures | 2/12 |  Neural Networks Foundations",5053,115000,1604793600,78,0,3,64],["Ilg3gGewQ5U","Backpropagation, intuitively | Deep Learning Chapter 3",767,6300000,1510099200,1,0,3,63],["PFDu9oVAE-g","Eigenvectors and eigenvalues | Chapter 14, Essence of linear algebra",1036,6400000,1478563200,1,4,3,63],["L8HKweZIOmg","Sequence-to-Sequence (seq2seq) Encoder-Decoder Neural Networks, Clearly Explained!!!",1010,387000,1699401600,3,0,3,63],["EkAQAi3a4js","Multiple Regression, Clearly Explained!!!",325,379000,1699401600,3,4,3,63],["HGwBXDKFk9I","Neural Networks Part 8: Image Classification with Convolutional Neural Networks (CNNs)",924,450000,1636329600,3,0,3,63],["68BZ5f7P94E","Neural Networks Pt. 3: ReLU In Action!!!",538,397000,1636329600,3,0,3,63],["wl1myxrtQHQ","The Chain Rule, Clearly Explained!!!",1104,420000,1604793600,3,4,3,63],["H3EjCKtlVog","Gaussian Naive Bayes, Clearly Explained!!!",566,452000,1604793600,3,0,3,63],["VX_M3tIyiYk","Power Analysis, Clearly Explained!!!",1005,413000,1604793600,3,4,3,63],["Rsc5znwR5FA","Statistical Power, Clearly Explained!!!",499,421000,1604793600,3,4,3,63],["Toet3EiSFcM","Support Vector Machines Part 2: The Polynomial Kernel (Part 2 of 3)",435,450000,1573171200,3,0,3,63],["8nm0G-1uJzA","Odds Ratios and Log(Odds Ratios), Clearly Explained!!!",980,428000,1541635200,3,0,3,63],["aeM-fmcdkXU","L8.8 Softmax Regression Derivatives for Gradient Descent",1180,17000,1636329600,5,0,3,63],["oY6-i2Ybin4","L6.2 Understanding Automatic Differentiation via Computation Graphs",1368,17000,1636329600,5,0,3,63],["htiNBPxcXgo","Lesson 8 - Practical Deep Learning for Coders 2022",5815,52000,1667865600,6,0,3,63],["WjnwWeGjZcM","Lesson 8 - Deep Learning for Coders (2020)",7018,47000,1604793600,6,0,3,63],["VEG5xT5gAHc","Lesson 7 - Deep Learning for Coders (2020)",7247,48000,1604793600,6,0,3,63],["AcA8HAYh7IE","Lesson 9 (2019) - How to train your model",8147,52000,1573171200,6,0,3,63],["0v93qHDqq_g","Intro to Machine Learning: Lesson 4",6005,48000,1541635200,6,0,3,63],["sHcLkfRrgoQ","Lesson 6: Deep Learning 2018",8023,54000,1541635200,6,0,3,63],["6kwQEBMandw","Lesson 3: Practical Deep Learning for Coders",7398,49000,1510099200,6,0,3,63],["4RPUrwzgO6c","Statistical Learning: 13.5 False Discovery Rate and Benjamini Hochberg Method",674,11000,1667865600,7,4,3,63],["Op0OyOuDjcQ","Statistical Learning: 9.1 Optimal Separating Hyperplane",696,12000,1667865600,7,0,3,63],["cdl4C2eCOHk","Statistical Learning: 8.5 Boosting",723,11000,1667865600,7,0,3,63],["_cKAxjnInfA","Statistical Learning: 8.4 Bagging",826,12000,1667865600,7,0,3,63],["2Zlcgt8FVz4","Stanford Seminar - Stopping grinding attacks in proofs of space",3963,12000,1541635200,7,5,3,63],["Dmm4UG-6jxA","MIT 6.S191 (2024): Deep Generative Modeling",3379,104000,1731024000,8,0,3,63],["AhyznRSDjw8","MIT 6.S191 (2023): Reinforcement Learning",3453,151000,1699401600,8,0,3,63],["AjtX1N_VT9E","MIT 6.S191 (2021): Convolutional Neural Networks",3357,164000,1636329600,8,0,3,63],["rZufA635dq4","MIT 6.S191 (2020): Deep Generative Modeling",2440,180000,1604793600,8,0,3,63],["o-_SZ_itxeA","Greedy? Min-p? Beam Search? How LLMs Actually Pick Words – Decoding Strategies Explained",713,8400,1762560000,10,0,3,63],["dOwRXpSSc8E","How do Vision Transformers work? – Paper explained | multi-head self-attention & convolutions",1155,20000,1667865600,10,0,3,63],["Nao16-6l6dQ","[Paper Analysis] The Free Transformer (and some Variational Autoencoder stuff)",2410,23000,1762560000,11,3,3,63],["AfAmwIP2ntY","Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters (Paper)",3182,31000,1731024000,11,3,3,63],["B45FlSQ8ITo","Scalable MatMul-free Language Modeling (Paper Explained)",2985,35000,1731024000,11,3,3,63],["hW3OVWfndLw","LLaMA Pro: Progressive LLaMA with Block Expansion (Paper Explained)",1905,35000,1731024000,11,3,3,63],["_okxGdHM5b8","Neural Networks are Decision Trees (w/ Alexander Mattick)",1911,58000,1667865600,11,0,3,63],["0PAiQ1jTN5k","How to make your CPU as fast as a GPU - Advances in Sparsity w/ Nir Shavit",3020,52000,1667865600,11,6,3,63],["rFwQDDbYTm4","[Classic] Playing Atari with Deep Reinforcement Learning (Paper Explained)",2352,53000,1604793600,11,3,3,63],["Nq3auVtvd9Q","[Classic] ImageNet Classification with Deep Convolutional Neural Networks (Paper Explained)",2767,52000,1604793600,11,3,3,63],["9suqiofCiwM","AutoGrad Changed Everything (Not Transformers) [Dr. Jeff Beck]",4598,22000,1767830400,13,0,3,63],["VZf6nHsIhso","Can LLMs Introspect? A Live Paper Review",8095,3700,1762560000,15,3,3,63],["IHikLL8ULa4","A Walkthrough of Progress Measures for Grokking via Mechanistic Interpretability: What? (Part 1/3)",2590,8400,1699401600,15,3,3,63],["sVo7SC62voA","SAM 3: The Eyes for AI  — Nikhila & Pengchuan (Meta Superintelligence), ft. Joseph Nelson (Roboflow)",4504,3700,1767830400,17,3,3,63],["xlXIi6GZNGY","⚡️Math Olympiad gold medalist explains OpenAI and Google DeepMind IMO Gold Performances",1987,4200,1762560000,17,4,3,63],["XTpyNrEgJQ4","Is Speculative Decoding Worth It? Profiling vLLM on NVIDIA Blackwell — Akamai",917,4800,1791417600,18,6,3,63],["YXowceUKYJI","KV Cache-Aware Routing and P/D Disaggregation on Kubernetes — Yuchen Fama & Ashish Kamra, Red Hat",1308,5100,1788825600,18,6,3,63],["6guQG_tGt0o","AI Kernel Generation: What's working, what's not, what's next – Natalie Serrino, Gimlet Labs",1155,7000,1767830400,18,6,3,63],["gFyBdBm0AGo","System Design for Next-Gen Frontier Models — Dylan Patel, SemiAnalysis",1109,6800,1762560000,18,6,3,63],["jGouxuAHIfQ","Lecture 101: Learning CUTLASS the hard way",4277,2800,1775606400,55,6,3,63],["sXDdRCy137c","Lecture 79 Mirage (MPK): Compiling LLMs into Mega Kernels",4372,3200,1762560000,55,6,3,63],["09wntC6BT5o","Lecture 9 Reductions",2802,4900,1731024000,55,6,3,63],["G5r2OyCN5_s","The AI Bandwidth Wall & Co-Packaged Optics",1034,161000,1762560000,56,6,3,63],["7p8ZeSbblec","The End of Dennard Scaling",931,160000,1762560000,56,6,3,63],["fc_xzN6UErI","Intel’s Next Breakthrough: Backside Power Delivery",1153,242000,1731024000,56,6,3,63],["yY8OFp0-UZw","The Crazy Computations Inside Your Smartphone Cameras",1048,353000,1667865600,56,5,3,63],["ckkrS752tjU","The Central Limit Theorem",657,37000,1762560000,57,4,3,63],["g-S0m2zcKUg","AI/ML+Physics Part 2: Curating Training Data [Physics Informed Machine Learning]",2171,60000,1731024000,57,0,3,63],["59huv1T_LJw","Complex Analysis L03: Functions of a complex variable, f(z)",1520,83000,1699401600,57,4,3,63],["9fGaTU1-f-0","Numerical Differentiation with Finite Difference Derivatives",2217,91000,1699401600,57,4,3,63],["ZSGrJBS_qtc","Eigenvalues and Eigenvectors",2037,95000,1667865600,57,4,3,63],["r1eWerqrcqo","Example Second-Order ODE: Spring-Mass-Damper",1980,93000,1667865600,57,4,3,63],["UXqUXYaRyGU","Deriving the Wave Equation",2156,103000,1667865600,57,4,3,63],["9d8PwnKVA-U","Deriving the Heat Equation: A Parabolic Partial Differential Equation for Heat Energy Conservation",1430,102000,1667865600,57,4,3,63],["MZJMi-6_4UU","Control Bootcamp:  Full-State Estimation",698,103000,1510099200,57,4,3,63],["SAk-6gVkio0","How CPU Memory & Caches Work - Computerphile",2094,138000,1762560000,59,5,3,63],["Ks1pw1X22y4","Random Numbers with LFSR (Linear Feedback Shift Register) - Computerphile",831,371000,1636329600,59,4,3,63],["7MT1F0O3_Yw","DNS Cache Poisoning - Computerphile",664,339000,1604793600,59,5,3,63],["gQddtTdmG_8","Vectoring Words (Word Embeddings) - Computerphile",1016,336000,1573171200,59,0,3,63],["rURRYI66E54","AI Language Models & Transformers - Computerphile",1239,354000,1573171200,59,0,3,63],["I5mRwzVvFGE","Spectre & Meltdown - Computerphile",825,358000,1541635200,59,5,3,63],["pLluFxHrc30","How WanaCrypt Encrypts Your Files - Computerphile",1042,373000,1510099200,59,5,3,63],["IAkj32VPcUE","Inside the CPU - Computerphile",676,376000,1510099200,59,5,3,63],["RPQD7-AOjMI","Turing Complete - Computerphile",386,366000,1478563200,59,5,3,63],["7ha78yWRDlE","Reverse Polish Notation and The Stack - Computerphile",812,327000,1415404800,59,5,3,63],["kgBjXUE_Nwc","Getting Sorted & Big O Notation - Computerphile",659,329000,1383868800,59,5,3,63],["v7yyZZjF1z4","[Unity] Procedural Cave Generation (E01. Cellular Automata)",1378,330000,1446940800,60,5,3,63],["ZymtNiyGB_g","Building the minutes and seconds for my clock: Breadboard clock part 4",1656,251000,1791417600,61,7,3,63],["0q6Ujn_zNH8","A simple BIOS for my breadboard computer",1313,519000,1731024000,61,5,3,63],["MCi7dCBhVpQ","SPI: The serial peripheral interface",1980,838000,1636329600,61,5,3,63],["DXvAlwMAxiA","How a transistor works",683,848000,1446940800,61,6,3,63]],"posts":[{"id":"2117655816581071203","p":37,"t":"There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing https://simonwillison.net/2026/Nov/3/qwen-can-interpret-model-weights/","ts":1793723465,"l":5424,"r":443,"c":412,"v":333346,"tp":3,"q":3,"s":79},{"id":"2112476496283775135","p":81,"t":"🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!\n\nPowering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.\n\n🔗 http://github.com/deepseek-ai/DeepInfer\n\n1/4","ts":1792488619,"l":43700,"r":4927,"c":1724,"v":7013003,"tp":6,"q":3,"s":82},{"id":"2118272105947037880","p":34,"t":"What a great idea! Learn LLM inference system on Apple Silicon for systems engineers: build a tiny vLLM + Qwen","ts":1793870400,"l":421,"r":29,"c":14,"v":18302,"tp":6,"q":3,"s":73,"qt":{"id":"2117929034177015929","n":"Oliver","h":"oliviscusAI","a":"https://pbs.twimg.com/profile_images/2008566513469427712/xbIOS-Fo_200x200.jpg","t":"Someone turned the hardest parts of LLM inference engineering into a course where you have to fix the performance problems yourself.\n\nit’s called tiny-llm, an open-source course for building a small vLLM-style inference engine around Qwen3.\n\nyou build a working model, then fix the bottlenecks that make it slow.\n\n> implement attention and text generation\n> add a KV cache to reuse previous calculations\n> work with packed 4-bit weights\n> write faster operations in C++ and Metal\n> handle multiple requests with batching and paged memory\n\nthe four-week course also covers a small coding agent with tool calls, checkpoints and context management.\neach stage includes exercises, reference solutions, tests and measurements.\n\ni like that you have to measure the improvements yourself.\n\nit uses MLX, C++ and Metal on Apple Silicon. No NVIDIA GPU required.\n\nthe introductory path supports Qwen3-0.6B on an 8GB Mac, while some later checkpoints remain under development.","ts":1793788605,"ph":[["https://pbs.twimg.com/media/HWRlLj2a4AA4U-S.png?name=orig",531,805]]}},{"id":"2118642061515936142","p":17,"t":"ok this is just ridiculous","ts":1793958604,"l":2515,"r":96,"c":143,"v":506040,"tp":8,"q":2,"s":72,"ph":[["https://pbs.twimg.com/media/HWbtufAacAAI8Qh.jpg?name=orig",690,984]],"qt":{"id":"2118489527362912566","n":"Thinking Machines","h":"thinkymachines","a":"https://pbs.twimg.com/profile_images/2075275500420001792/Q9J4M1DL_200x200.jpg","t":"Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.\n\nInkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.\n\nhttps://thinkingmachines.ai/news/inkling-2-small","ts":1793922237}},{"id":"2118993073455698264","p":22,"t":"recommended reading.\n\nhttps://www.anthropic.com/research/components-of-agent-memory","ts":1794042292,"l":952,"r":72,"c":24,"v":58823,"tp":2,"q":3,"s":71},{"id":"2114054277379641770","p":85,"t":"vLLM matches DeepInfer, DeepSeek's own engine for DeepSeek-V4.1-Flash, on DeepSeek's decode configuration across batch sizes 8 to 256, and does it at DEP32 against DeepInfer’s EP64DP64.\n\nDeepInfer was originally built on vLLM. By integrating DeepSeek's open-source kernels and extensively optimizing prefill and decode, vLLM now matches DeepInfer's performance curves.","ts":1792864791,"l":1099,"r":121,"c":58,"v":124926,"tp":6,"q":3,"s":78,"ph":[["https://pbs.twimg.com/media/HVahKXZa4AINwCv.jpg?name=orig",1920,1080]]},{"id":"2116305528330785190","p":17,"t":"its true, Anthropic really vibecoded their entire backend to Rust and none of you noticed","ts":1793401531,"l":3285,"r":160,"c":235,"v":304281,"tp":2,"q":2,"s":74,"qt":{"id":"2116050219989373170","n":"Somi","h":"somi_ai","a":"https://pbs.twimg.com/profile_images/1937022514137374720/-y9Y42z2_200x200.jpg","t":"Anthropic rewrote its Messages API in Rust. Zero Rust devs on the team.\n\nClaude wrote the code. Humans wrote the plan and checked old vs new matched on real traffic. CPU dropped 40-50%.\n\nPretty sure \"never do the big rewrite\" is dead advice.\n\n@katelyn_lesse on @latentspacepod","ts":1793340661,"vd":["https://video.twimg.com/amplify_video/2116050170949599232/vid/avc1/1920x1080/0EvjC48GStPVWr3a.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2116050170949599232/img/h-K1sSC2lJwkXBHy.jpg",1920,1080]}},{"id":"2118403114890137895","p":76,"t":"Introducing Personal Finance in Claude, in beta in the US.\n\nConnect the accounts you choose, and Claude answers your money questions in everyday language, from your own balances and transactions. Access is read-only, and Claude can't move, manage, or invest your money.","ts":1793901635,"l":45432,"r":2657,"c":1482,"v":10579454,"tp":8,"q":2,"s":70,"vd":["https://video.twimg.com/amplify_video/2118401044841390080/vid/avc1/1920x1080/lQ99Zn7JR9hyYGA6.mp4?tag=29","https://pbs.twimg.com/media/HWYUk7ZWwAAcrVe.jpg",3841,2160]},{"id":"2117312276328075644","p":23,"t":"I made a wiki to learn RL, enjoy: https://0xsero.github.io/home-rl-wiki/wiki/learning-path","ts":1793641559,"l":958,"r":46,"c":25,"v":28435,"tp":0,"q":3,"s":72,"ph":[["https://pbs.twimg.com/media/HWI0OSeaEAAELL4.jpg?name=orig",3080,1966]]},{"id":"2102435511222890900","p":76,"t":"Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family.\n\nIt performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.","ts":1790094661,"l":98354,"r":9074,"c":3367,"v":29639989,"tp":8,"q":3,"s":84,"vd":["https://video.twimg.com/amplify_video/2102432417352929280/vid/avc1/1920x1080/yilx3O9YWu7w9vaX.mp4?tag=29","https://pbs.twimg.com/media/HS1aPgYWMAAP9sw.jpg",1280,720]},{"id":"2117375052618539485","p":15,"t":"This is a cool paper, but I think it's best thought of as Astra is surprisingly capable of doing old school mechanistic interpretability research rather than anything spooky. The things it's doing are clever, but typically exploiting structure in the data and using known method mechanistic interpretability techniques (which Astra was surely pretrained on), though it is better at this than I would have predicted!\n\nThis is an exciting development for the prospects of ambitiously scaling interpretability or doing very laborious analysis like this, on more real models. However, I'm not holding my breath. This kind of analysis was pretty common up until about 2022, it largely fell out of favor because a lot of it is just exploiting fairly clean surface-level patterns that doesn't scale to the annoying complexity of real LLMs or which only works for the more simpler and superficial properties. And as far as I can tell, basically all the case studies here are just using a bag of interpretability tricks rather than anything spooky. The piece doesn't really update me on how far old school mech interp can go, just on how good LLMs are at it. It's very fun to see the creative solutions Astra came up with though!","ts":1793656526,"l":1860,"r":164,"c":53,"v":104093,"tp":3,"q":3,"s":71,"qt":{"id":"2117313742891327994","n":"Rikiya Takehi","h":"rikiyatakehi","a":"https://pbs.twimg.com/profile_images/1731888727897567232/97PMSyds_200x200.jpg","t":"We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.\nIt directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.\n\nWe call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵\nBlog: https://rikiyat.github.io/blog/weight-space-programming/","ts":1793641908,"ph":[["https://pbs.twimg.com/media/HWI1oKUWUAAWoZp.jpg?name=orig",1200,625]]}},{"id":"2112703311933747347","p":85,"t":"🚨🚨🚨 vLLM is all you need to reproduce DeepSeek’s performance.\n\nWith Wide-EP and EPLB, vLLM can match DeepInfer’s decode performance within 6% and generate $39/gpu/hr in revenue at DeepSeek’s off-peak pricing, at 200 P90 TPS on the AgentX workload.\n\nSpecifically:\n• Under DeepSeek’s decode configuration (DeepSeek-V4.1-Flash, 131K context, nextn=3), vLLM on GB200 achieves throughput within 6% of DeepInfer’s published results across batch sizes 8–256.\n• On AgentX, vLLM pushes the Pareto frontier further on GB300, reaching up to 1M tokens/s/GPU and generating $38.96/gpu/hr in revenue at 196 P90 TPS, based on DeepSeek’s off-peak API pricing.\n• Also, vLLM extended low-latency kernel support for DeepSeek-V4.1-Flash, achieving up to 826 TPS per user on AgentX.","ts":1792542696,"l":1040,"r":97,"c":36,"v":114258,"tp":6,"q":3,"s":77,"ph":[["https://pbs.twimg.com/media/HVHSmVGbAAA8RBY.jpg?name=orig",1664,1452]]},{"id":"2112588257104216294","p":38,"t":"New blog post 🎉: How To Use Decision Models For Evals\n\nIn this post, we show you how to choose a decision model for your task, tune its threshold (don't assume the 0.5 default is good!), and interpret its scores.\n\nLots of cool interactive charts here. https://x.com/i/article/2112552453422104576","ts":1792515265,"l":2319,"r":259,"c":198,"v":326276,"tp":2,"q":3,"s":77},{"id":"2118519423455932882","p":23,"t":"First American Open Weight model that mogs all the Chinese Open weight models \n\nVery impressive, inkling is my fav US lab.\n\nLooking forward to optimizing this","ts":1793929365,"l":2280,"r":75,"c":44,"v":126439,"tp":1,"q":2,"s":68,"qt":{"id":"2118489527362912566","n":"Thinking Machines","h":"thinkymachines","a":"https://pbs.twimg.com/profile_images/2075275500420001792/Q9J4M1DL_200x200.jpg","t":"Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.\n\nInkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.\n\nhttps://thinkingmachines.ai/news/inkling-2-small","ts":1793922237}},{"id":"2118026522179317874","p":85,"t":"🎉 Kimi K3.1 Pro from @Kimi_Moonshot lands in vLLM on day 0!\n\nA 1M-token context window for long coding sessions, 3D creation and research. It scores 65.7% on Terminal-Bench 4.0 with the Claude Code harness.\n\nGive Kimi the long task. vLLM brings Day-0 serving across GPUs and TPUs, plus RL support.","ts":1793811848,"l":263,"r":29,"c":11,"v":16322,"tp":8,"q":3,"s":69,"vd":["https://video.twimg.com/amplify_video/2118026494199136256/vid/avc1/1920x1080/07AK_rlkWqHxwNYW.mp4?tag=16","https://pbs.twimg.com/amplify_video_thumb/2118026494199136256/img/IeWdYL5OMq984KH7.jpg",1920,1080],"qt":{"id":"2118010528123773297","n":"Kimi.ai","h":"Kimi_Moonshot","a":"https://pbs.twimg.com/profile_images/2086719038798102529/diJGKpOE_200x200.jpg","t":"Releasing the model weights of Kimi K3.1 Pro: https://huggingface.co/moonshotai/Kimi-K3.1-Pro\nWe'll keep advancing open intelligence.","ts":1793808035,"ph":[["https://pbs.twimg.com/media/HWSvVQBaMAAOxi5.jpg?name=orig",3840,2160]]}},{"id":"2118814753154064501","p":75,"t":"Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. https://www.anthropic.com/research/components-of-agent-memory","ts":1793999777,"l":11873,"r":1048,"c":849,"v":1471392,"tp":2,"q":3,"s":67},{"id":"2117706239010594848","p":23,"t":"Omarchy Linux on AppleSilicon gives you more memory bandwidth than MacOS","ts":1793735487,"l":567,"r":13,"c":25,"v":23088,"tp":6,"q":3,"s":69,"qt":{"id":"2117528278663897439","n":"Joshua Warren","h":"JoshuaSWarren","a":"https://pbs.twimg.com/profile_images/2095930967525302273/HVcHM0x-_200x200.jpg","t":"Linux now beats macOS on Apple silicon GPU memory bandwidth.\n\nM1 Max, Omarchy:\nmacOS: 361 GB/s\nLinux before: 241 GB/s\nLinux now: 368 GB/s\n\nThe kernel never told the power controller the GPU was busy. One bit fixed it.\n\nhttps://github.com/omacom/linux-aurora/pull/334","ts":1793693058}},{"id":"2104633115620823187","p":76,"t":"Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family.\n\nIt’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.","ts":1790618611,"l":55631,"r":4440,"c":1951,"v":12782410,"tp":8,"q":3,"s":81,"vd":["https://video.twimg.com/amplify_video/2104629272543399936/vid/avc1/1920x1080/psIdtRJzfm_fmopA.mp4?tag=29","https://pbs.twimg.com/media/HTUthyeWIAAInna.jpg",1920,1080]},{"id":"2115114028338962672","p":78,"t":"This is Genie 4, our most realistic and interactive world model yet. 🌐\n\nExplore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.","ts":1793117455,"l":23856,"r":2638,"c":1026,"v":3517849,"tp":8,"q":2,"s":73,"vd":["https://video.twimg.com/amplify_video/2115110795553964032/vid/avc1/1920x1080/UjdIAfbWqSTv6JaM.mp4?tag=29","https://pbs.twimg.com/media/HVplvENXUAA-sCS.jpg",1920,1080]},{"id":"2116997870213906524","p":22,"t":"recommended reading. had the good fortune to talk with @tianyi a little while ago, and the DSH folks are one of the not so many teams doing innovative stuff in the harness space.\n\nthe below largely aligns with what we are gravitating towards at earendil as well. single, long lived entity, ad-hoc, freeform, possibly recursive delegation for actual tasks, cross-system event ingestion and triggers.\n\ngood stuff.","ts":1793566598,"l":784,"r":44,"c":19,"v":101890,"tp":2,"q":3,"s":70,"qt":{"id":"2116857257245987236","n":"DeepSeek Harness","h":"DeepSeekHarness","a":"https://pbs.twimg.com/profile_images/2105141374123495424/_om7PZyR_200x200.jpg","t":"https://x.com/i/article/2116829218935857152","ts":1793533074}},{"id":"2118449808402681911","p":14,"t":"I spent 20,000 dollars in API credits having different AI models play each other at Blood on the Clocktower (a Mafia or Werewolf type social deception game).\n\nThe results were fascinating. Models have gotten much better over the course of this year at creative on-the-spot scheming and bluffing, and also at reasoning as a team to break apart these underground conspiracies.\n\nI go through some of the most standout games these AIs played in this video. The new models are getting pretty good.  If my friends pulled off some of these moves while we were playing, I’d give them a fist bump.\n\nFwiw, OpenAI’s newer models are much better than Anthropic’s at this game.\n\nYou can see the full benchmark at the website, and I’ve also made it so that you and your friends can play again the AIs: http://clocktowerbench.com","ts":1793912767,"l":2955,"r":170,"c":180,"v":491288,"tp":3,"q":3,"s":67,"vd":["https://video.twimg.com/amplify_video/2118447221465100289/vid/avc1/1920x1080/F4XOvVzCanOvIe0S.mp4?tag=29","https://pbs.twimg.com/media/HWZY5oDWIAAiAi4.jpg",1920,1080]},{"id":"2117445039211086096","p":39,"t":"This is a fantastic idea - a mass attempted replication of old findings of LLMs with new models. Much conventional wisdom is outdated.\n\nRecent models no longer invent citations and both hallucination rates & stereotyping has dropped. But they are still sycophantic & lean left.","ts":1793673212,"l":1698,"r":157,"c":234,"v":107190,"tp":3,"q":3,"s":69,"ph":[["https://pbs.twimg.com/media/HWKsoQxWAAAScMH.jpg?name=orig",1222,1490],["https://pbs.twimg.com/media/HWKs-I1WsAAl-_N.jpg?name=orig",1441,1817],["https://pbs.twimg.com/media/HWKs_brXUAA0sdn.jpg?name=orig",1462,1784],["https://pbs.twimg.com/media/HWKtBt6WYAEtJh9.jpg?name=orig",1471,1924]]},{"id":"2107894039626277339","p":76,"t":"Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we’ve ever released.\n\nOn average, it costs around 75% less to run than Claude Haiku 4.5.","ts":1791396076,"l":45172,"r":3254,"c":1610,"v":6808342,"tp":8,"q":3,"s":79,"vd":["https://video.twimg.com/amplify_video/2107890134603481088/vid/avc1/3840x2160/KakPq_aXsoFB9L23.mp4?tag=29","https://pbs.twimg.com/media/HUC-jfVW8AAz_Ic.jpg",1920,1080]},{"id":"2094848572143407483","p":76,"t":"We’re introducing Claude Fable 5.1 and Claude Mythos 5.1.\n\nThey're the world’s most advanced models for coding and knowledge work.","ts":1788285794,"l":65237,"r":7272,"c":2859,"v":21130893,"tp":8,"q":3,"s":82,"vd":["https://video.twimg.com/amplify_video/2094843261470793728/vid/avc1/2160x2160/GEnONyeW_6cMSzmW.mp4?tag=29","https://pbs.twimg.com/media/HRJlwmVWcAEyYzf.jpg",1620,1620]},{"id":"2118848233762078971","p":43,"t":"libghostty 1.0 is here! I started Ghostty ~5yrs ago to work towards this. Today, libghostty is in hundreds of projects used by tens of millions of people. 1.0 marks ABI stability, quality, and an invite for distros to begin packaging. https://mitchellh.com/writing/libghostty-1-0\n\nSince starting the Ghostty project in 2022, the most common question I've been asked is \"why?\". Today, I'll answer this by showing the broad industry impact libghostty has already had.\n\nOn top of JetBrains and Neovim already mentioned (that alone being over 10 million users):\n-KDE developers are experimenting with a Konsole replacement built on it. \n- Vercel's Turborepo uses it to render task output in its TUI. \n- devenv uses it to drive shells. \n- Antithesis used it to fuzz TUI programs and found real bugs in btop and others. \n-Ghostel brings it to Emacs. \n- ghostty-web is the WASM build with an Xterm.js-compatible API. \n- Commercial products like OrbStack embed it. \n- And a whole wave of new terminals and \"agentic\" coding tools, such as herdr and cmux, have been built on it rather than writing yet another emulator.\n\nThere are community bindings for Rust, Go, JavaScript, Swift, Python, .NET, Elixir, Dart, Java, and more, and the awesome-libghostty list tracks well over a hundred projects.\n\nMore people today use a terminal powered by libghostty than use Ghostty the application, by a very large multiple. And libghostty is having a much larger impact than being \"just\" a terminal emulator.\n\nI'm excited to see where people take it from here.","ts":1794007759,"l":4228,"r":271,"c":96,"v":89174,"tp":5,"q":3,"s":65},{"id":"2118362983285850288","p":45,"t":"I regularly hear from people that my post-training book/writing has helped them get a job at labs by focusing on material that develops intuitions for how training methods work. It's very validating, because this type of \"idea/math heavy\" writing doesn't really sell many books. It's rewarding the right people who put the effort into learning. I'm happy I made the time for it.","ts":1793892067,"l":2058,"r":64,"c":60,"v":86278,"tp":0,"q":2,"s":66},{"id":"2118359553532481655","p":45,"t":"Have heard from many folks that score-centering is popular in industry (e.g. Tinker's StableReinforce using it), so it's the latest addition to the rlhf book. It's a nice derivation to understand training-inference-mismatch.\n\nVideo version explaining it soon.","ts":1793891249,"l":425,"r":28,"c":17,"v":22299,"tp":0,"q":3,"s":66,"qt":{"id":"2100997113903477121","n":"Martin Marek","h":"mrtnm","a":"https://pbs.twimg.com/profile_images/2103063091537723392/X98MIkgY_200x200.jpg","t":"Why is RL so sensitive to training-inference mismatch (TIM)?\n\nWith each training step, the trainer is pulled toward the sampler, resulting in accumulating drift.\n\nWe exploit this intuition to devise a novel correction method to stabilize RL under TIM, called Score Centering.","ts":1789751721,"vd":["https://video.twimg.com/tweet_video/HSg9E15XUAA-nxF.mp4","https://pbs.twimg.com/tweet_video_thumb/HSg9E15XUAA-nxF.jpg",1200,632]}},{"id":"2117338841749029221","p":39,"t":"Waiting for replication but this is a pretty big deal.","ts":1793647892,"l":1473,"r":73,"c":101,"v":75876,"tp":3,"q":3,"s":68,"qt":{"id":"2117313742891327994","n":"Rikiya Takehi","h":"rikiyatakehi","a":"https://pbs.twimg.com/profile_images/1731888727897567232/97PMSyds_200x200.jpg","t":"We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.\nIt directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.\n\nWe call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵\nBlog: https://rikiyat.github.io/blog/weight-space-programming/","ts":1793641908,"ph":[["https://pbs.twimg.com/media/HWI1oKUWUAAWoZp.jpg?name=orig",1200,625]]}},{"id":"2047516922263285776","p":81,"t":"🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.\n\n🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.\n🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.\n\nTry it now at http://chat.deepseek.com via Expert Mode / Instant Mode. API is updated & available today!\n\n📄 Tech Report: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf\n🤗 Open Weights: https://huggingface.co/collections/deepseek-ai/deepseek-v4\n\n1/n","ts":1777001049,"l":45342,"r":7433,"c":1649,"v":10172285,"tp":8,"q":3,"s":82,"ph":[["https://pbs.twimg.com/media/HGo9ronbUAAKlRk.jpg?name=orig",2809,554]]},{"id":"2117637894315626719","p":5,"t":"Catching up on the 10 (!) open LLM architecture releases from the past month. Quite a lot happened in October. \nI guess we can call it \"Open October\"!\n\n1. Kolibri-1 (78B): A notable model from Germany. Features classic Grouped Query Attention (GQA) and Sliding Window Attention (SWA) attention module.\n\n2. Ling 3.1 Flash (560B): Uses Kimi Delta Attention and the eighth uses gated Multi-head Latent Attention (MLA) inspired by DeepSeek.\n\n3. Muse Spark Open (816B): Besides Musse Glimmer, the first major open weight release from Meta (since Llama). Uses classic GQA + SWA but adds a LatentMoE.\n\n4. Nemotron 3.5 Super VL (120B): Combines Mamba-2 layers with just eight attention layers. Also LatentMoE \n\n5. Dory1.0 (4B): A looped transformer design.\n\n6. Reflection Beam (501B): GQA+SWA architecture.\n\n7. Recursive Flint (40B): Note that \"Recursive\" is the company name; the architecture is a fairly standard MoE with GQA and SWA.\n\n8. MiniMax M3.1-Flash (428B): Uses their own MiniMax Sparse Attention module.\n\n9. LFM3 (1.7B): Hybrid architecture that mixes short convolutions and attention.\n\n10. Mistral Large 4 (1T) combines MLA with sparse attention.\n\nWill probably cover/discuss things in more depth in the coming weeks!","ts":1793719192,"l":1637,"r":224,"c":106,"v":63226,"tp":0,"q":3,"s":67,"ph":[["https://pbs.twimg.com/media/HWNcY1KXYAAC0Ge.jpg?name=orig",4096,4096]]},{"id":"2118119451401048497","p":77,"t":"Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.\n\nIn a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API cost of $1.2 million.","ts":1793834004,"l":12798,"r":1337,"c":679,"v":2260418,"tp":3,"q":3,"s":66,"ph":[["https://pbs.twimg.com/media/HWUQl7QbwAAkznq.png?name=orig",3840,2160]]},{"id":"2117042619662701016","p":23,"t":"Broke 1000 tok/s prefill on an M4 Mac Mini (16gb ram/256gb ssd) + RTX 3090 \n\nThe cost to run one of the best models ever is like 2200","ts":1793577267,"l":2132,"r":64,"c":97,"v":121063,"tp":1,"q":2,"s":68,"ph":[["https://pbs.twimg.com/media/HWE-3Q4bkAApRfB.jpg?name=orig",2966,1646]]},{"id":"2118073650356384239","p":17,"t":"> we developed and trained Ripple-65B Preview for only $32M, which is one-tenth the cost of the previous open-weight SOTA on agentic coding¹\n\nthis is like the avengers of open model SOTA\n\n¹ \"Note that $32M is our full R&D cost starting from scratch, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run.\"","ts":1793823084,"l":213,"r":11,"c":123,"v":23232,"tp":1,"q":3,"s":66,"ph":[["https://pbs.twimg.com/media/HWTopt3bMAA8AGV.jpg?name=orig",721,1029]],"qt":{"id":"2118046765857546733","n":"Ricursive Intelligence","h":"RicursiveAI","a":"https://pbs.twimg.com/profile_images/1995674420178980864/7dWOHGam_200x200.jpg","t":"Today we release Ripple-65B Preview, our 65B dense LLM we pretrained from scratch.\n\nRipple sets a new state-of-the-art for U.S. open-weight models in agentic coding tasks, surpassing significantly larger models including Muse Spark Open-816B, Beam-501B, Inkling-975B, and Nemotron 3 Ultra-550B.\n\nLearn more: https://www.ricursive.com/ripple","ts":1793816675,"vd":["https://video.twimg.com/amplify_video/2118013573670793217/vid/avc1/1920x1080/hc-3oimWhcv_Lhw3.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118013573670793217/img/RNjAFnyHuXndpo2P.jpg",1920,1080]}},{"id":"2084100707423289643","p":80,"t":"📢Meet Qwen3.8-Max — our most capable model to date. \n\nNext week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!🎉\n\nQwen3.8-Max, a new bar for coding and cowork at 2.4T parameters:\n\n- Autonomous coding: 10+ days of self-evolving development, from empty folder to production without hand-holding, complete project trace in the GitHub：https://github.com/qwen-code-dev-bot/oh-my-cli\n- Real work, real results: Production-quality deliverables across hundreds of professions.\n- Long-horizon mastery: System-level autonomous planning with closed-loop adaptive learning, driving 500+ turns of chip design optimization and 365 days of e-commerce strategy.\n- Native multimodal intelligence: Vision isn't just input — it's a continuous feedback loop for planning, execution, and self-correction.\n\n💰Pricing:\nInput: $2.0 / M tokens\nOutput: $6.0 / M tokens\nImplicit Caching: $0.25 / M tokens\n\nStart building with Qwen3.8-Max! 🚀\n\n📖 Blog: https://qwen.ai/blog?id=qwen3.8\n✅ Qwen Studio: https://chat.qwen.ai/?models=qwen3.8-max\n⚡ API: https://www.qwencloud.com/models/qwen3.8-max","ts":1785723304,"l":22529,"r":2780,"c":1226,"v":7265060,"tp":8,"q":3,"s":81,"ph":[["https://pbs.twimg.com/media/HOw1F2ebgAAkpJJ.jpg?name=orig",4096,3014]]},{"id":"2097930608790167907","p":81,"t":"🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.\n\n🔹 Introducing the smallest model in our new architecture family, with native visual understanding.\n🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models.\n\n1/6","ts":1789020609,"l":28762,"r":3078,"c":1107,"v":7205329,"tp":8,"q":3,"s":80,"ph":[["https://pbs.twimg.com/media/HR1UyHiaAAAtpqw.jpg?name=orig",4096,2101]]},{"id":"2115489356437651654","p":78,"t":"Gemini 4 Carbon is our next frontier model.\n\nIt’s state-of-the-art across cybersecurity and long-horizon capabilities.","ts":1793206941,"l":14504,"r":1388,"c":751,"v":1685251,"tp":8,"q":2,"s":70,"ph":[["https://pbs.twimg.com/media/HVu5JfhXkAAHYVr.jpg?name=orig",4096,3447]]},{"id":"2115444742595260551","p":37,"t":"I think October represents another inflection point in capabilities for coding agents and the models that drive them: Claude Opus 5.5 can now build video games that are actually fun https://simonwillison.net/2026/Oct/28/vibe-game-development/","ts":1793196304,"l":1376,"r":76,"c":414,"v":124954,"tp":2,"q":3,"s":70},{"id":"1881318130334814301","p":81,"t":"🚀 DeepSeek-R1 is here!\n\n⚡ Performance on par with OpenAI-o1\n📖 Fully open-source model & technical report\n🏆 MIT licensed: Distill & commercialize freely!\n\n🌐 Website & API are live now! Try DeepThink at http://chat.deepseek.com today!\n\n🐋 1/n","ts":1737376170,"l":34985,"r":6660,"c":1937,"v":12566223,"tp":8,"q":3,"s":81,"ph":[["https://pbs.twimg.com/media/GhvI3AoaAAAw-4z.png?name=orig",3821,2793]]},{"id":"1846591870165799017","p":57,"t":"New 20hr bootcamp on Probability & Statistics!!!\n\nVideos released weekly but full playlist already posted:\nhttps://www.youtube.com/watch?v=sQqniayndb4&list=PLMrJAkhIeNNR3sNYvfgiKgcStwuPSts9V&index=1\n\nProbability & Statistics are cornerstones of data science and machine learning. This course rapidly covers the basics and gets into advanced topics.","ts":1729096784,"l":9511,"r":1517,"c":76,"v":535619,"tp":4,"q":3,"s":81,"vd":["https://video.twimg.com/ext_tw_video/1846591438584414208/pu/vid/avc1/1280x720/q8hWzhAYKc6z5ZJP.mp4?tag=12","https://pbs.twimg.com/ext_tw_video_thumb/1846591438584414208/pu/img/igweyii_2yhkAILU.jpg",1280,720]},{"id":"2118755687274532893","p":41,"t":"so far this is working great. i tried this in the past and it wasn't as great. some details\n\n- we have long running sessions per project\n- opus is told to do all work in deepseek subagents\n- be strict about scope when creating a task\n- ask for proof of work instead of a summary\n- do a one step review of the diff to make sure it's not totally off\n\nopus is amazing at this role","ts":1793985695,"l":1850,"r":41,"c":100,"v":81982,"tp":2,"q":3,"s":64,"qt":{"id":"2118059907949690955","n":"dax","h":"thdxr","a":"https://pbs.twimg.com/profile_images/1602333093485891584/mmVqjFNI_200x200.jpg","t":"the team is experimenting this week at switching all our subagents to deepseek flash and using opus 5.5 only on the primary subagent","ts":1793819808}},{"id":"2118755418675765656","p":78,"t":"We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption.\n\nGuided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.\n\nOur VP Science @Pushmeet explains more ↓ https://x.com/pushmeet/status/2118752095859794390","ts":1793985631,"l":1343,"r":178,"c":270,"v":116023,"tp":3,"q":3,"s":64,"qt":{"id":"2118752095859794390","n":"Pushmeet Kohli","h":"pushmeet","a":"https://pbs.twimg.com/profile_images/650718639784509440/yaRuXASm_200x200.jpg","t":"https://x.com/i/article/2118746079923437568","ts":1793984838}},{"id":"2114696948301152573","p":45,"t":"My first open model lecture, covering all the many pieces that build the ecosystem, and the open questions/debates determining where it goes next.\n\nHow does distillation fit in?\nHow do we manage risks?\nHow do we see access to models changing in the future?\n\nI'm excited to revisit these questions about every 6 months, to give a state of the union on the field.\n\nChapters\n00:00 Introduction\n01:36 What is an open model?\n08:47 Why should we care about open models?\n18:52 How competitive are the best open models today?\n25:28 How are open models built? The role of distillation\n34:52 Are people actually using open models?\n41:47 What comes next with open models?\n47:32 Conclusions\n\nWhat questions did I not answer that you have?","ts":1793018016,"l":1083,"r":129,"c":62,"v":59212,"tp":1,"q":3,"s":71,"vd":["https://video.twimg.com/amplify_video/2114694763366490112/vid/avc1/1920x1080/rBD4TG0FEcaDqRXm.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2114694763366490112/img/YHHBKiXuXCI8Uv12.jpg",1920,1080]},{"id":"2113688040912863492","p":24,"t":"listen to me very carefully","ts":1792777474,"l":3893,"r":114,"c":47,"v":235738,"tp":7,"q":2,"s":72,"qt":{"id":"2113684128734625948","n":"Crystal","h":"crystalsssup","a":"https://pbs.twimg.com/profile_images/2017077042395959307/Weh0_Alq_200x200.jpg","t":"Kimi K3.1 Pro is very good at 3D modeling.\n\nI asked Kimi K3.1 to create a 3D website that pulls apart a robotic hand into 2,317 pieces. It did pretty well!\n\nYou can just build things.\n\nProject link: https://form-01-score.ok.kimi.link/","ts":1792776541,"vd":["https://video.twimg.com/amplify_video/2113684075009777664/vid/avc1/2318x1440/ZXvVmA0qCh0t-bY7.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2113684075009777664/img/R8f7H5lOCpTRXeaR.jpg",2318,1440]}},{"id":"2112760256120197333","p":85,"t":"Quantization is easy. Fast quantization kernels for every format × every GPU × every workload are not.\n\nToday we're open-sourcing Humming with Ant Group: one kernel library for all of it — up to 5x kernel speedups, +14–43% E2E MoE throughput.\n\nBlog + benchmarks: https://vllm.ai/blog/2026-10-21-humming","ts":1792556272,"l":488,"r":47,"c":25,"v":35302,"tp":6,"q":3,"s":73},{"id":"2111401488392343722","p":85,"t":"🎉 Congrats @deepseek_ai on DeepSeek-V4.1-Flash-1017! Better scores and shorter outputs make it a great fit for long agent loops. Making each step fast is the serving side's job.\n\nOn DeepSeek-V4.1-Flash, vLLM delivers 1.3B tokens per dollar TCO on SemiAnalysis AgentX: 830K tok/s per GPU (cached + input + output) at 100 tok/s per user. 🚀\n\nIt's been a fun few weeks building with the community. Chapter one, 5x agentic throughput since day 0:\n🔗 https://vllm.ai/blog/2026-10-07-deepseek-v41-flash\n\nNext chapter coming soon. 👀","ts":1792232317,"l":616,"r":57,"c":29,"v":56133,"tp":6,"q":3,"s":74,"ph":[["https://pbs.twimg.com/media/HU0zOFJaIAAF9Wc.jpg?name=orig",1846,1602]],"qt":{"id":"2111375640717033777","n":"DeepSeek","h":"deepseek_ai","a":"https://pbs.twimg.com/profile_images/1717417613775757312/Uk1zNOj4_200x200.jpg","t":"🚀 DeepSeek-V4.1-Flash-1017 is here!\n\nStronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.\n\n🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.\n\n1/4","ts":1792226154,"ph":[["https://pbs.twimg.com/media/HU0b65KbsAAypqr.jpg?name=orig",1052,889],["https://pbs.twimg.com/media/HU0b7UfbwAAGXNH.jpg?name=orig",2937,2531]]}},{"id":"2117524701354361269","p":46,"t":"I guess 'tis the day i complain about over-hype on this app.\n\nThis does not show a quadrillion param model, nor how to train one.\n\nThe claim (which is only in the blog and this thread, the paper is ok) is based on wild extrapolations, but i promise you that the techniques studied in the paper will not continue carrying more than ~one OOM.\n\nIn fact you can are already see here (figure) that reality starts detaching from prediction. It will continue detaching more.\n\n(The study is on 25M-400M unique tokens, 30M-400M params)\n\nAlso, just because i can't resist sprinkling a little bit of Schmidhubering, how to do many epochs is something vision folks studied a long time, and the \"increase wd with epochs\" is a finding were published in S4L paper in 2019. It's of course worth revisiting in LLMs but i generally find things carry over.","ts":1793692205,"l":564,"r":29,"c":19,"v":43883,"tp":3,"q":3,"s":66,"ph":[["https://pbs.twimg.com/media/HWL1fanWkAAHTxQ.jpg?name=orig",1280,600]],"qt":{"id":"2117370524330594698","n":"Samip","h":"industriaalist","a":"https://pbs.twimg.com/profile_images/2033799858679910404/EPdMVEea_200x200.jpg","t":"Humans are ~1,000,000x more sample efficient than LLMs, and there are no good explanations.\n\nWe’ve found that a big part of the answer is that LLMs are *massively* underparameterized by many orders of magnitude. The brain has 100T synapses learning from ~200M tokens of language. Chinchilla prescribes a 1T-parameter model on 20T tokens. That’s 7 OOMs apart in parameters per token.\n\nAfter a lot of painstaking work, we present new data-constrained scaling laws and learning algorithms to scale to (extrapolated) quadrillion-parameter transformers.\n\nw/ @akshayvegesna, @EthanBaron78553, @asatheesh6561\nFull paper led by @EthanBaron78553, with @ShmuelBerman, @andrewgwils","ts":1793655446,"ph":[["https://pbs.twimg.com/media/HWJpRUvbAAAEJOy.jpg?name=orig",1200,1091]]}},{"id":"2112480719952138301","p":85,"t":"Grateful to @deepseek_ai for releasing DeepInfer, initially built primarily on @vllm_project, under Apache 2.0 and crediting the open-source community!\n\nWe've worked hard on DeepSeek V4.1 Flash inference efficiency and will keep improving with ideas from DeepInfer! Stay tuned 🙌","ts":1792489626,"l":2384,"r":134,"c":58,"v":83329,"tp":6,"q":2,"s":73,"ph":[["https://pbs.twimg.com/media/HVEJ3G7XoAE-vSF.png?name=orig",834,221]],"qt":{"id":"2112476496283775135","n":"DeepSeek","h":"deepseek_ai","a":"https://pbs.twimg.com/profile_images/1717417613775757312/Uk1zNOj4_200x200.jpg","t":"🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!\n\nPowering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.\n\n🔗 http://github.com/deepseek-ai/DeepInfer\n\n1/4","ts":1792488619}},{"id":"2118859591798980657","p":39,"t":"Some new data from Anthropic looking at the impact of AI on the sciences by field: \"With current capabilities, if adopted widely, AI could take on 27–52% of the researcher time needed to produce a new study, with a median of 38% across fields...\"\nHere: https://www.anthropic.com/research/how-researchers-use-claude","ts":1794010467,"l":653,"r":92,"c":181,"v":45421,"tp":3,"q":3,"s":63,"ph":[["https://pbs.twimg.com/media/HWextZ8WcAE5veF.jpg?name=orig",1920,2450],["https://pbs.twimg.com/media/HWezgLFXYAAL7DW.jpg?name=orig",1920,1658]]},{"id":"2118380813498892301","p":47,"t":"The era of System One Vision Model is here! Introducing VFlash: parsing the visual world in parallel at unmatched speeds for Physical AI.\n\nMost embodied prediction tasks like object localization, robot trajectory planning, and dense 3D motion are fixed-length float vectors. For example, a bounding box is always 6 tokens, a 10-waypoint robot arm trajectory is 54: we know these beforehand. Instead of decoding one token at a time, we can output the entire geometric shape in one go. \n\nVFlash enables up to 12x faster action decoding and a 5.1x overall robot inference speedup on edge GPUs like the Jetson Thor and DGX Spark. What’s surprising is that it ends up even more accurate than the VLM backbone, with up to 61% lower 3D trajectory error. The intuition is that VFlash avoids the cascading errors of autoregressive decoding, where every coordinate is written in ink and every later point loses the chance to backspace.\n\nEverything that moves will be autonomous, and everything autonomous will need to think on its own chip. The next inference revolution is physical reasoning at the speed of reflex.\n\nAnd as always, it’s open paper, open source! Link in thread:","ts":1793896318,"l":1311,"r":130,"c":139,"v":66915,"tp":3,"q":3,"s":64,"vd":["https://video.twimg.com/amplify_video/2118374869645864960/vid/avc1/3840x2160/4KKdRrgyqBnLezX6.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118374869645864960/img/X0zFDyMQe_jbJcnE.jpg",3840,2160]},{"id":"2117353378925940987","p":23,"t":"You can have LLMs literally modify and rewrite other models weights directly.\n\nI had this little side project to try and get an LLM to reconstruct and entire 270m paran model from memory","ts":1793651358,"l":347,"r":15,"c":34,"v":12512,"tp":3,"q":3,"s":66,"qt":{"id":"2117313742891327994","n":"Rikiya Takehi","h":"rikiyatakehi","a":"https://pbs.twimg.com/profile_images/1731888727897567232/97PMSyds_200x200.jpg","t":"We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.\nIt directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.\n\nWe call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵\nBlog: https://rikiyat.github.io/blog/weight-space-programming/","ts":1793641908,"ph":[["https://pbs.twimg.com/media/HWI1oKUWUAAWoZp.jpg?name=orig",1200,625]]}},{"id":"2097374640582668336","p":77,"t":"We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.\n\nThe proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.\n\nThe problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.","ts":1788888056,"l":119838,"r":20011,"c":5658,"v":75401898,"tp":4,"q":3,"s":79,"ph":[["https://pbs.twimg.com/media/HRtS_iLboAUUlYv.jpg?name=orig",1254,1254]]},{"id":"2112984664256241837","p":76,"t":"We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent.\n\nIt comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.","ts":1792609775,"l":55559,"r":4266,"c":2428,"v":11678757,"tp":8,"q":2,"s":72,"vd":["https://video.twimg.com/amplify_video/2112982862186106880/vid/avc1/640x360/L7zosiEUW76tqYr-.mp4?tag=29","https://pbs.twimg.com/media/HVLUnMDWsAAn7Pw.png",1920,1080]},{"id":"2116653884362097081","p":22,"t":"that will do, slopus. that will do.\n\nintroducing pi-schema. a vibe+hill-climbed json schema builder and validator library that works in any JS env, including anything without eval(). just a silly closure \"jit\".\n\nfast, correct, low mem and bundle size footprint.","ts":1793484586,"l":524,"r":11,"c":28,"v":65544,"tp":2,"q":3,"s":67,"ph":[["https://pbs.twimg.com/media/HV_c-nOXwAATHWa.png?name=orig",2400,1012],["https://pbs.twimg.com/media/HV_dIllXUAA-kGA.png?name=orig",2400,2056],["https://pbs.twimg.com/media/HV_dL_RX0AEq0W9.png?name=orig",2400,1012],["https://pbs.twimg.com/media/HV_dPVCXAAAr27b.png?name=orig",2400,1012]],"qt":{"id":"2116633270377160975","n":"Mario Zechner","h":"badlogicgames","a":"https://pbs.twimg.com/profile_images/1553485821767991296/87k3l720_200x200.jpg","t":"you know what's a perfect hill climbing target? a json schema generator and validator library, faster than anything under the (node) sun, so i can finally get right of typebox everywhere.","ts":1793479671}},{"id":"2105152056382345544","p":82,"t":"Ollama now supports Jev-like decision models all locally.\n\nUse decision models like Nimble for tasks like ticket triaging, model routing, and content moderation.\n\nollama pull nimble\n\nHere’s Nimble playing Ollama racer through the new local /v1/systemone API by making decisions in real-time. 🏎️","ts":1790742336,"l":4592,"r":605,"c":174,"v":667908,"tp":1,"q":3,"s":77,"vd":["https://video.twimg.com/amplify_video/2105148364274995200/vid/avc1/3360x2160/2ON9PV9BJzkS7csT.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2105148364274995200/img/JeWi_6c-GQgEeM9e.jpg",3360,2160]},{"id":"2113683001519284270","p":45,"t":"TLDR on the kimi k3.1 pro benchmarks is that it looks like they figured out their RL, where the initial release was a very impressive base model with lighter post-training. China is very much keeping pace with capability gains.","ts":1792776272,"l":1066,"r":40,"c":33,"v":49493,"tp":8,"q":3,"s":71},{"id":"2114315936953925952","p":45,"t":"Nice thread, some RL details used by closed labs slowly leaking out.","ts":1792927176,"l":867,"r":33,"c":15,"v":55202,"tp":0,"q":3,"s":70,"qt":{"id":"2114145574987702525","n":"Martin Marek","h":"mrtnm","a":"https://pbs.twimg.com/profile_images/2103063091537723392/X98MIkgY_200x200.jpg","t":"Tinker's new StableReinforce objective is Reinforce + Score Centering. Here's how I reverse-engineered it 🧵","ts":1792886558}},{"id":"2112509913847210471","p":45,"t":"This is really a wild way to get adoption on your models.\nReally great for the community, and the timing is hilarious given the govt just banned deepseek in federal agencies.","ts":1792496586,"l":1168,"r":50,"c":40,"v":62519,"tp":6,"q":3,"s":72,"qt":{"id":"2112476496283775135","n":"DeepSeek","h":"deepseek_ai","a":"https://pbs.twimg.com/profile_images/1717417613775757312/Uk1zNOj4_200x200.jpg","t":"🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!\n\nPowering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.\n\n🔗 http://github.com/deepseek-ai/DeepInfer\n\n1/4","ts":1792488619}},{"id":"2088280182356611304","p":80,"t":"We promised open weights for Qwen3.8. Now, time to meet them! 🎉\n\n⚡ Qwen3.8-27B:\n- A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows.\n- 262K native context, easily extendable to 1M tokens via YaRN.\n- Built for builders. Highly efficient, high-quality, and licensed under Apache 2.0.\n\n🚀 The open weights for Qwen3.8-2.4T-A95B (Max-level) have also been released recently.\n\nWhether you're shipping lightweight applications with Qwen3.8-27B locally or building agents with Qwen3.8-2.4T-A95B, they're yours now!\n\nDownload, deploy, and build something we haven't imagined yet. 👀👇\n- Hugging Face：\nhttps://huggingface.co/collections/Qwen/qwen38\n- ModelScope：\nhttps://www.modelscope.cn/collections/Qwen/Qwen38","ts":1786719768,"l":15896,"r":2110,"c":849,"v":6049058,"tp":8,"q":3,"s":79,"ph":[["https://pbs.twimg.com/media/HPsKxzsbQAAJa4d.jpg?name=orig",2048,1152]]},{"id":"2118405434071810210","p":34,"t":"Magical idea!","ts":1793902188,"l":73,"r":1,"c":0,"v":8680,"tp":3,"q":3,"s":63,"qt":{"id":"2117808859691811216","n":"Jarrod Watts","h":"jarrodwatts","a":"https://pbs.twimg.com/profile_images/2046035436435648512/OLuqKfK9_200x200.jpg","t":"I asked Astra to edit Qwen3.8-27B's weights and make it more token efficient.\n\nFor ~$100 in GPU cost, it found the weights for the \"done thinking\" signal and edited them to be twice as strong.\n\nThe result: 40% fewer thinking tokens on average on the pelican on a bicycle test:","ts":1793759953,"ph":[["https://pbs.twimg.com/media/HWP2vT6a4AAGz5E.jpg?name=orig",2400,1350]]}},{"id":"2105388084154056939","p":78,"t":"Introducing Gemini 4 Argon – our new frontier model.\n\nIt’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.","ts":1790798610,"l":42728,"r":4619,"c":1982,"v":9046644,"tp":8,"q":2,"s":76,"ph":[["https://pbs.twimg.com/media/HTfV52SWgAIlX3W.png?name=orig",984,987]]},{"id":"2111814711956168760","p":85,"t":"🚀 vLLM v0.32.0 is out! 814 commits · 363 contributors · 121 first-timers\n\n🧭 New /v1/systemone structured-decision endpoint + generic hf tool/reasoning parser\n🗜️ Low-precision KV everywhere: NVFP4 on SM8x/90/12x, FP8 KV default for GLM-5.x, 4-bit KV on ROCm\n🛠️ Model Runner V2: GPU n-gram speculator, Elastic EP, PP+CP for MLA, now default on ROCm\n⚡ Faster startup & serving: precompiled kernel downloads, autotune preload, +43–85% req/s with multi-API workers\n🆕 EmbeddingGemma 2, MiniCPM-V 4.7, Nemotron 3.5 ASR\n🔒 3 security fixes + CORS hardening\n\nFull notes → http://github.com/vllm-project/vllm/releases/tag/v0.32.0","ts":1792330837,"l":422,"r":45,"c":42,"v":24513,"tp":6,"q":3,"s":72,"ph":[["https://pbs.twimg.com/media/HU6sLh-bQAAMRVN.jpg?name=orig",1600,900]]},{"id":"2116602645972672576","p":41,"t":"hex was fully vibe coded. the app. the landing page. the animations. sound effects were ai generated. issues people open are fixed by opencode. the onboarding. even some of the creative ideas\n\nbut i think it's pretty good software. it's fun to use. works on all platforms. performant (not electron!)\n\nour effort didn't go into the code. but there was a lot of effort\n\nit's a small app but i hope it improves your outlook on where we can go","ts":1793472370,"l":2697,"r":39,"c":188,"v":177672,"tp":2,"q":3,"s":66},{"id":"2116570248933691801","p":23,"t":"They made a model into a computer \n\nI can’t wait till this catches on. You give the model circuits for calculations","ts":1793464645,"l":367,"r":13,"c":18,"v":20366,"tp":0,"q":3,"s":66,"qt":{"id":"2115394340357116169","n":"div (fp32)","h":"divagr1925","a":"https://pbs.twimg.com/profile_images/2103115932990332929/wKj91D1E_200x200.jpg","t":"1/4\n\nMeet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM.\n\nIt runs C and MicroPython in its forward passes, no external runtime.\n\nOn DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.","ts":1793184287,"vd":["https://video.twimg.com/amplify_video/2115390746090237952/vid/avc1/1280x720/re3aWl-8X1nw-xGd.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2115390746090237952/img/O2d_O9rEssR_l5qF.jpg",1280,720]}},{"id":"2116527005801873662","p":41,"t":"with current token prices i wouldn't really say \"cost of software is going to zero\"\n\nour token cost is double our payroll which means even though we're 25 it's like we're 75\n\nand i feel like our output is similar to a team of 75 pre-ai. we're just able to do it without all the downsides of a larger team\n\nthis is still a good thing. we couldn't hire 50 more people. so much overhead and impossible to retain a high bar","ts":1793454336,"l":2895,"r":92,"c":207,"v":188078,"tp":2,"q":3,"s":66},{"id":"2118037508139413742","p":32,"t":"We're releasing new Unsloth Dynamic v3.1 GGUFs for Qwen3.8-Flash-Next! They're better by 10-15% accuracy on DeepSWE, Terminal Bench, Aider Polyglot and more vs GSQ-RCO.\n\nWe also provide a new engram switching guide for BF16, Q8_0, Q5_1, IQ4_NL!\n\nThis builds upon our DeepSeek-V4.1-Flash UD3.1 work, which improved over 10-20% accuracy vs others.\n\nAll benchmarks are 256K for agentic ones, and 128K for the rest. We used temperature 1.0, top_p 0.95, top_k 20, min_p 0, no presence or repetition penalty. preserve_thinking and xhigh reasoning, and ran over 26B of tokens for benchmarking.\n\nThe RMSE vs BF16 for each of the engrams:\nBF16   102.4 GB    0\nQ8_0     54.4 GB   0.0053\nQ5_1     38.4 GB   0.0376\nIQ4_NL  28.8 GB   0.0757\n\nMore details at https://unsloth.ai/docs/models/qwen3.8-next#dynamic-v3.1-update-november-2nd-2026. We also moved all benchmarking, grading, infra issues and more to https://unsloth.ai/docs/basics/dynamic-3.0-ggufs/dynamic-3.1-ggufs","ts":1793814467,"l":196,"r":19,"c":11,"v":9409,"tp":1,"q":3,"s":63,"ph":[["https://pbs.twimg.com/media/HWTG_1EakAAiIPd.jpg?name=orig",2400,1380]],"qt":{"id":"2118033433045991660","n":"Unsloth AI","h":"UnslothAI","a":"https://pbs.twimg.com/profile_images/2036313340616777728/OLALWB2__200x200.jpg","t":"We’re releasing new Qwen3.8-Flash-Next GGUFs with 15% higher accuracy.\n\nUnsloth Dynamic V3.1 outperforms others by >15% on Terminal Bench, DeepSWE & more benchmarks (256K context).\n\nRun on 32GB+ RAM setups.\n\nGGUF: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF\nBlog: https://unsloth.ai/docs/models/qwen3.8-next","ts":1793813496,"ph":[["https://pbs.twimg.com/media/HWTEDzGboAAl8mB.jpg?name=orig",3724,4096]]}},{"id":"2111375640717033777","p":81,"t":"🚀 DeepSeek-V4.1-Flash-1017 is here!\n\nStronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.\n\n🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.\n\n1/4","ts":1792226154,"l":37231,"r":4131,"c":1629,"v":8378520,"tp":8,"q":2,"s":72,"ph":[["https://pbs.twimg.com/media/HU0b65KbsAAypqr.jpg?name=orig",1052,889],["https://pbs.twimg.com/media/HU0b7UfbwAAGXNH.jpg?name=orig",2937,2531]]},{"id":"2116379267504152848","p":41,"t":"guys deepseek flash is already natively quantized. the experts are FP4\n\nthere's little point in quantizing it further you barely get anything out of it\n\nwhy do so many people say \"they're quantizing deepseek\" this doesn't make any sense to me","ts":1793419112,"l":2946,"r":46,"c":139,"v":148597,"tp":1,"q":3,"s":66},{"id":"2118497681995661338","p":45,"t":"Looks like the smartest western open weight model by a margin. While much smaller than muse spark open or mistal large. While with a better license (apache). Love it!","ts":1793924181,"l":942,"r":41,"c":24,"v":52809,"tp":1,"q":2,"s":62,"qt":{"id":"2118489527362912566","n":"Thinking Machines","h":"thinkymachines","a":"https://pbs.twimg.com/profile_images/2075275500420001792/Q9J4M1DL_200x200.jpg","t":"Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.\n\nInkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.\n\nhttps://thinkingmachines.ai/news/inkling-2-small","ts":1793922237}},{"id":"2117477032879874539","p":23,"t":"If you have a few idle GPUs give this to Claude/codex and ask it to implement this on Qwen 27B \n\nI know for a fact that transplanting attention algos onto different models is possible","ts":1793680840,"l":1177,"r":64,"c":31,"v":63139,"tp":3,"q":2,"s":64,"qt":{"id":"2117317159777177602","n":"James Whittington","h":"jcrwhittington","a":"https://pbs.twimg.com/profile_images/2087267816923676672/-60kzVYS_200x200.jpg","t":"Today we are happy to share one of our first open research releases: Latent Sparse Attention (LSA), a step towards subquadratic attention that matches accuracy of full attention at ultra-long context.\n\nSparse attention attends to only a few tokens per query, but it has to find the right ones. At 10M tokens, that search takes 98% of the attention layer's time. LSA makes sparse attention up to 22.7× faster and full attention 72.4x faster.\n\nWith @kubicek_ales @caciolai @PengLiangzu @SukritSumant @Ben_Hoov @FraserGreenlee @RMBattleday @radim_urban @thoefler @jcrwhittington","ts":1793642723,"vd":["https://video.twimg.com/amplify_video/2117305488186441728/vid/avc1/1920x1080/C0VdNbSixMI-AX2b.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2117305488186441728/img/fhf-8Nh8dVdmpp02.jpg",1920,1080]}},{"id":"2116271281351881120","p":45,"t":"Was reading the score-centering RL paper to add some notes on it the my book, and they have a great table in the appendix on the different importance sampling based train-inference mismatch methods in RL. These have been the biggest focus of RL algorithm research since the GRPO takeoff, imo.","ts":1793393366,"l":458,"r":34,"c":33,"v":26927,"tp":0,"q":3,"s":66,"ph":[["https://pbs.twimg.com/media/HV6Bg_2WQAA5p6R.jpg?name=orig",1428,1264]]},{"id":"2114135647040778667","p":85,"t":"Beyond matching DeepInfer's performance, we also put vLLM to the test on @SemiAnalysis_ AgentX with real-world agentic workloads.\n\nAt DeepSeek's off-peak pricing, vLLM delivers $38.96/GPU-hour revenue, with 954K tokens/s/GPU total throughput (cached+input+output) at 196 tokens/s P90 per user.\n\nThis is the best DeepSeek-V4.1-Flash result on AgentX ever, powered by recent advancement in vLLM!\n\nCheck out our blog for the full optimization story and how to reproduce the results 👇 Enjoy the read.","ts":1792884191,"l":276,"r":33,"c":26,"v":40209,"tp":6,"q":3,"s":69,"ph":[["https://pbs.twimg.com/media/HVbrKsdaoAAicF9.jpg?name=orig",1846,1484]],"qt":{"id":"2114054289488478256","n":"vLLM","h":"vllm_project","a":"https://pbs.twimg.com/profile_images/1774187681746182144/N_5NJ8B1_200x200.jpg","t":"Scaled out with P/D disaggregation and Wide-EP on a GB300 NVL72 rack, vLLM serves V4.1-Flash on SemiAnalysis AgentX at 954K tokens/s per GPU with 196 tokens/s P90 per user, 7.9× our single-node result.\n\nOur engineering blog covers the work behind it: https://vllm.ai/blog/2026-10-23-deepseek-v41-flash-pd-wide-ep","ts":1792864794}},{"id":"2117355349967544808","p":22,"t":"recommended reading. i have many feelings about this.\n\nhttps://rikiyat.github.io/blog/weight-space-programming/","ts":1793651828,"l":1305,"r":95,"c":33,"v":113439,"tp":3,"q":2,"s":64},{"id":"2116773135668404286","p":17,"t":"btw @karpathy uses medium on opus 5.5 on his latest project (Hackertube launched today).\n\noverall i think having strong opinions on effort levels is like the tabs vs spaces of our time. just tweak to taste. altho mine is definitely superior to yours.","ts":1793513017,"l":824,"r":29,"c":278,"v":83074,"tp":2,"q":2,"s":65,"ph":[["https://pbs.twimg.com/media/HWBJ6PNbEAAHuO6.jpg?name=orig",2124,1884],["https://pbs.twimg.com/media/HWBJ8i0bQAA2-A5.jpg?name=orig",2054,1832]],"qt":{"id":"2116642586387730593","n":"swyx 🔜 aie.codes","h":"swyx","a":"https://pbs.twimg.com/profile_images/2073162797354217472/hNny55eF_200x200.jpg","t":"join me all the way on the light side guys\n\nlow effort + critique loop > superultramaxhighwhatever\n\n(for the throughput-constrained; see swyxio/skills/babysit-runs)","ts":1793481892}},{"id":"1911746525698379842","p":3,"t":"I do my best to tell exiting tales...here's my latest about Reinforcement Learning with Neural Networks: Mathematical Details!!!\nhttps://youtu.be/DVGmsnxB2UQ","ts":1744630864,"l":594,"r":53,"c":5,"v":58160,"tp":0,"q":3,"s":78},{"id":"1892926638893007260","p":3,"t":"THIS IS FREEEEEEEEEE!!! And, if I say so myself, completely awesome. BAM!","ts":1740143854,"l":605,"r":46,"c":10,"v":25796,"tp":0,"q":3,"s":78,"qt":{"id":"1892655862835249449","n":"DeepLearning.AI","h":"DeepLearningAI","a":"https://pbs.twimg.com/profile_images/2099541018089607169/6_ec0Cx__200x200.jpg","t":"Last week, we launched \"Attention in Transformers: Concepts and Code in PyTorch\" instructed by @JoshuaStarmer! In this course, you'll:\n\n✅ Learn how the attention mechanism in LLMs helps convert base token embeddings into rich context-aware embeddings.\n\n✅ Understand the Query, Key, and Value matrices, what they are for, how to produce them, and how to use them in attention.\n\n✅ Learn the difference between self-attention, masked self-attention, and cross-attention, and how multi-head attention scales the algorithm.\n\n🔗 Enroll for free: https://hubs.la/Q037ytWB0","ts":1740079296,"vd":["https://video.twimg.com/ext_tw_video/1892650448206680064/pu/vid/avc1/1280x720/JYFLVXUqhRspBD-L.mp4?tag=12","https://pbs.twimg.com/ext_tw_video_thumb/1892650448206680064/pu/img/yrhqcItM45pN3Y3x.jpg",1920,1080]}},{"id":"1758861944377708948","p":10,"t":"We simply explain and illustrate Mamba and (Selective) State Space Models – SSMs.\n📺 https://youtu.be/vrF3MtGwD0Y\nSSMs match performance of transformers, but are faster and more memory-efficient than them. This is crucial for long sequences!\nIncredible work by @_albertgu @tri_dao! 👏","ts":1708180340,"l":236,"r":53,"c":2,"v":21269,"tp":0,"q":3,"s":78,"ph":[["https://pbs.twimg.com/media/GGi6htVXIAADFoM.jpg?name=orig",1008,562]]},{"id":"2112960942078107960","p":45,"t":"Finally got to put together a complete deck on the state of open models.\nWhat they are, licenses, how competitive they are, signs people are using them, distillation's impact, and how we balance the risks and benefits of openness!\nWill record a video lecture of these soon, and hopefully re-do them a few times a year.","ts":1792604120,"l":783,"r":90,"c":34,"v":39443,"tp":1,"q":3,"s":70,"ph":[["https://pbs.twimg.com/media/HVK-x4tWwAAI2LF.jpg?name=orig",3272,1834]]},{"id":"2095595741528125780","p":77,"t":"This is GPT-6 Astra.\n\nAnything you can do on a computer, Astra can do for you. Fast.","ts":1788463933,"l":338111,"r":35148,"c":9120,"v":139263539,"tp":8,"q":2,"s":77,"vd":["https://video.twimg.com/amplify_video/2095595661559574528/vid/avc1/1920x1080/grXdGA07X-yPH_8z.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2095595661559574528/img/Vmb2pgEFJ6fpCUTD.jpg",1920,1080]},{"id":"2111637219647840611","p":85,"t":"Kimi K3 decode on NVIDIA GB300 NVL72: 842 tokens/s at batch 1 with DSpark 🚀\n\nA unified CuTe DSL decode path fuses attention, MoE and communication, delivering 1.5–2× higher decode throughput in several tested configurations and up to 1.8× higher P90 interactivity on the @SemiAnalysis_ AgentX workload.\n\nThe post covers the kernel work, the correctness and long-context evals behind it, and benchmark sweeps across batch sizes, context lengths and parallelism configurations.\n\nFull write-up: http://vllm.ai/blog/2026-10-16-kimi-k3-megakernel\n\nGreat work and collaboration with the @NVIDIA team!","ts":1792288520,"l":356,"r":38,"c":23,"v":73445,"tp":6,"q":3,"s":71},{"id":"2113522286061924746","p":46,"t":"On the zeroth order lm training stuff:\n\nIt's cool that people try stuff instead of only reading and dooming. But i feel the need to curb the enthusiasm on this one a little. Two things:\n\n(1) methods in the genetic-algo and fwd-diff categories, which imo these fall in, is established that they are fine for smaller parameter counts but fall off hard as parameter space increases. It's not necessarily raw param count that matters here but \"intrinsic\" size. This is the whole point of backprop: few outputs many params.\n\n(2) loss 4.2 for LM is very high even for a 30M parameter model. Go a bit higher and you are at learning token unigram distribution. This means for this score not much interesting/difficult learning has to happen yet.\n\nAgain it's fine to rediscover things like this from first principles. In fact it's the very best way to learn and i recommend it; i spent a lot of time learning like this myself! But now with social media hype amplification and crazy vcs and journalist attention that our field gets, i do feel the need to be a bit more cautious with communication and there to be \"old pessimist\" voice like mine in the room to.\n(But unlike some other old pessimist voices, if i turn out to be wrong I'll happily admit it and cheer on =)","ts":1792737954,"l":953,"r":39,"c":31,"v":54253,"tp":3,"q":3,"s":69,"qt":{"id":"2113375116528877668","n":"spicylemonade","h":"spicey_lemonade","a":"https://pbs.twimg.com/profile_images/2043453291196002304/q0yHemFP_200x200.jpg","t":"WE DID IT! We can now pretrain transformers without backprop and even beat backprop+Adam on validation loss with insane speeds.\n\nNow the big question is scale. It’s Ironic, we have reached a breakthrough In AI capability due to inspiration from a safety blog","ts":1792702867}},{"id":"2113483068174692831","p":39,"t":"By every measure, Opus 4.8 beats human experts as peer reviewers in a study of 124 Harvard social science papers covering a range from statistical analyses to formal models to qualitative work.\n\nAs academics, we need to figure out how to adapt the practice of science to AI, fast.","ts":1792728604,"l":1719,"r":261,"c":260,"v":107818,"tp":3,"q":3,"s":69,"ph":[["https://pbs.twimg.com/media/HVSZLTTWIAAvpHM.png?name=orig",839,664],["https://pbs.twimg.com/media/HVSZYYfXMAEH3DN.jpg?name=orig",845,1095]]},{"id":"2114325097653669940","p":5,"t":"The complete reasoning-from-scratch video series is now up (and linked) here: https://magazine.sebastianraschka.com/p/building-reasoning-models\n\n(Taking a short break but will be back with LLM model and research updates in November!)","ts":1792929360,"l":1772,"r":257,"c":85,"v":107223,"tp":0,"q":3,"s":68},{"id":"2116987671939875319","p":39,"t":"GPT-4 could do 4 minutes of independent work with a 50% chance of success.\n\nCurrent models can do multiple weeks of work, and January 2026 is a reasonable cut-off for the moment agentic work started.\n\nLarge productivity impacts before that would mostly be confined to coding.","ts":1793564167,"l":774,"r":41,"c":260,"v":44156,"tp":3,"q":3,"s":64,"qt":{"id":"2116967827773030602","n":"Alex Imas","h":"alexolegimas","a":"https://pbs.twimg.com/profile_images/1994816403359010817/0zMDGpbF_200x200.jpg","t":"Steve was very early on this. But he’s exactly right—chatbot era was never going to be a huge productivity boost. Analyses should put the line at January 2026; agentic AI is where productivity gains are.","ts":1793559436}},{"id":"2118033433045991660","p":84,"t":"We’re releasing new Qwen3.8-Flash-Next GGUFs with 15% higher accuracy.\n\nUnsloth Dynamic V3.1 outperforms others by >15% on Terminal Bench, DeepSWE & more benchmarks (256K context).\n\nRun on 32GB+ RAM setups.\n\nGGUF: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF\nBlog: https://unsloth.ai/docs/models/qwen3.8-next","ts":1793813496,"l":2186,"r":216,"c":124,"v":240802,"tp":1,"q":3,"s":62,"ph":[["https://pbs.twimg.com/media/HWTEDzGboAAl8mB.jpg?name=orig",3724,4096]]},{"id":"2110047081414987957","p":75,"t":"New on our Engineering blog: Creating in-house source control. \n\nIn May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.  \n\nHere, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: https://www.anthropic.com/engineering/redesigning-source-control","ts":1791909401,"l":25966,"r":2662,"c":1174,"v":8361175,"tp":5,"q":3,"s":72},{"id":"2087887408440164663","p":81,"t":"🧩 DeepSeek Harness v0.1 is now available in Developer Preview!\n\n🔹 We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT license.\n🔹 Powered by the Cordis meta-framework, DeepSeek Harness is an agent harness built around one core idea: Everything is a plugin. Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are ALL implemented as plugins, and can be mixed, matched, replaced, and extended.\n\nTry it now!\nhttps://github.com/deepseek-ai/deepseek-harness","ts":1786626123,"l":20040,"r":2437,"c":845,"v":4506951,"tp":2,"q":3,"s":77},{"id":"2098097512544444447","p":75,"t":"We're publishing our most detailed threat intelligence report to date. \n\nIt covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.\n\nWe disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies.\n\nThese cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve.\n\nWe’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop.\n\nRead the report: https://www.anthropic.com/threat-intelligence-report-september-2026","ts":1789060402,"l":50208,"r":11643,"c":3203,"v":43907749,"tp":3,"q":3,"s":76},{"id":"2102824959827742916","p":75,"t":"Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR.\n\nWe don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA. Historically, the discovery of such programmable systems has helped revolutionize medicine. CRISPR, for instance, is now the foundation of genetic medicines. But it will take much more work to learn what this system does, and whether it can be put to similar use.\n\nRead more: https://www.anthropic.com/news/claude-discovers-novel-enzyme-system","ts":1790187513,"l":40725,"r":5279,"c":1552,"v":25701472,"tp":3,"q":3,"s":75},{"id":"2075581603443511663","p":27,"t":"QuadRF is a phased array antenna that can see WiFi through walls, spot drones in the sky... and it's powered by a Pi 5 and FPGA board. Very neat device, we have a full video up on Geerling Engineering...","ts":1783692191,"l":3941,"r":432,"c":86,"v":173086,"tp":7,"q":3,"s":77,"vd":["https://video.twimg.com/amplify_video/2075581323784093696/vid/avc1/1080x1920/15v033zDggQhJpQ-.mp4?tag=28","https://pbs.twimg.com/amplify_video_thumb/2075581323784093696/img/GYhoH68ZMWuJZQ_7.jpg",1080,1920]},{"id":"2116723717829230932","p":39,"t":"This is both neat & a sign of why it is proving to be so hard to isolate agents from each other. They share a common \"brain\" which means they don't have a lot of trouble recognizing creative communication solutions generated by other AIs\n\nWonder if it works across model families?","ts":1793501235,"l":705,"r":57,"c":203,"v":55380,"tp":3,"q":3,"s":64,"qt":{"id":"2116704239003349503","n":"Robocurve PBC","h":"robocurve","a":"https://pbs.twimg.com/profile_images/2093086017268097024/qtVm8mqA_200x200.jpg","t":"We asked two isolated copies of the same AI to exchange a secret word with only a robot and a camera.\n\nThey succeeded by using air-writing, Morse code, flag semaphore and a 5-bit binary alphabet, without agreeing on a protocol beforehand.","ts":1793496591,"vd":["https://video.twimg.com/amplify_video/2116701680326852608/vid/avc1/1920x1080/b32KzHpfDDg3tJtM.mp4?tag=29","https://pbs.twimg.com/media/HWAZQvMawAAnJGn.jpg",1920,1080]}},{"id":"2033167146302210058","p":5,"t":"I (finally) put together a new LLM Architecture Gallery that collects the architecture figures all in one place!\nhttps://sebastianraschka.com/llm-architecture-gallery/","ts":1773579796,"l":8091,"r":1429,"c":198,"v":740340,"tp":0,"q":3,"s":77,"ph":[["https://pbs.twimg.com/media/HDdCsMqWAAAML2O.jpg?name=orig",3276,2808]]},{"id":"1986828899591020603","p":55,"t":"The most intuitive explanation of floats I've ever come across, courtesy of @fabynou https://fabiensanglard.net/floating_point_visually_explained/","ts":1762531897,"l":1563,"r":175,"c":17,"v":148637,"tp":0,"q":3,"s":77,"ph":[["https://pbs.twimg.com/media/G5KiU0faYAAyHhD.jpg?name=orig",1423,955]]},{"id":"1889714317974442331","p":3,"t":"My FREE @DeepLearningAI course on Attention is LIVE!!! \nhttps://bit.ly/4hDKXYg","ts":1739377977,"l":486,"r":46,"c":17,"v":44704,"tp":0,"q":3,"s":77,"qt":{"id":"1889713576782582228","n":"DeepLearning.AI","h":"DeepLearningAI","a":"https://pbs.twimg.com/profile_images/2099541018089607169/6_ec0Cx__200x200.jpg","t":"Last week we introduced how transformer LLMs work, this week we go deeper into one of its key elements—the attention mechanism, in a new course, Attention in Transformers: Concepts and Code in PyTorch, taught by 🪕🎶 @joshuastarmer 🪕🎶 founder of StatQuest!\n\nTransformers, the architecture behind models like ChatGPT, rely on attention to process language efficiently. Instead of treating words in isolation, they weigh relationships between them, creating richer, context-aware representations.\n\nIn this new course, Josh breaks down the core ideas behind attention and transformers, step by step. You’ll explore: \n\n🔹 How attention evolved from early NLP models to transformers\n🔹 The role of Query, Key, and Value matrices\n🔹 Self-attention, masked self-attention, and cross-attention\n🔹 How multi-head attention improves performance\n\nEnroll in this course. For free. Double Bam! https://bit.ly/3EwueI3","ts":1739377800,"vd":["https://video.twimg.com/ext_tw_video/1889695414930071552/pu/vid/avc1/1280x720/rvfBagDF13nlWG4U.mp4?tag=12","https://pbs.twimg.com/ext_tw_video_thumb/1889695414930071552/pu/img/L7qoErR0-Fq603Xf.jpg",1920,1080]}},{"id":"1889679823003333091","p":3,"t":"My FREE @DeepLearningAI  course on Attention is LIVE!!! Learn the concepts and pytorch code behind the most fundamental part of the AI revolution!\nhttps://bit.ly/4hDKXYg","ts":1739369753,"l":478,"r":68,"c":7,"v":21856,"tp":0,"q":3,"s":77},{"id":"1738180641366175765","p":10,"t":"Finetuning LLMs is now possible without reinforcement learning, but with Direct Preference Optimization (DPO). \n📺 https://youtu.be/XZLc09hkMwA\nIn this video, we explain:\n👉the DPO paper\n👉RLHF and why reinforcement learning was needed in the first place.","ts":1703249533,"l":209,"r":48,"c":3,"v":27692,"tp":0,"q":3,"s":77,"ph":[["https://pbs.twimg.com/media/GB9BDiuXcAAR-bq.jpg?name=orig",1008,562]],"qt":{"id":"1734375040932831537","n":"Chelsea Finn","h":"chelseabfinn","a":"https://pbs.twimg.com/profile_images/1531702062487965698/a6iO2pRx_200x200.jpg","t":"DPO is a runner up for NeurIPS outstanding paper. 🙌\n\nBig congrats especially to the students @rm_rafailov @archit_sharma97 @ericmitchellai & the other awardees.\n\nIf you haven't learned about DPO already, check out the oral & poster 👇 on Thurs afternoon.","ts":1702342207,"ph":[["https://pbs.twimg.com/media/GBGxPDTa0AEyjUd.jpg?name=orig",2646,1760]]}},{"id":"2102460975790137662","p":77,"t":"Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.\n\nGPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.\n\nWe’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.","ts":1790100733,"l":53125,"r":5242,"c":2221,"v":10220141,"tp":8,"q":3,"s":75,"vd":["https://video.twimg.com/amplify_video/2102460948430966784/vid/avc1/1920x1080/2aqucgwb6tQ7xZ6U.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2102460948430966784/img/AtB4MxYZuAKgOLey.jpg",1920,1080]},{"id":"2107596713791767021","p":77,"t":"We’re releasing a broad range of new mathematical results produced by an internal frontier model.\n\nWe’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. \n\nhttps://github.com/openai/math","ts":1791325188,"l":40626,"r":6289,"c":1901,"v":22788264,"tp":4,"q":3,"s":73},{"id":"2117710942826123294","p":75,"t":"New Anthropic research: progress on provable inference for large language models.\n\nHere, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. \n\nThese proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.\n\nFor our full write-up, see here: https://www.anthropic.com/research/provable-inference\n\nFurther progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: https://github.com/anthropics/FlashProve\n\nWe’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (http://rdalal.github.io) and Ben Mayer (@transmissions11).","ts":1793736608,"l":5252,"r":541,"c":650,"v":700880,"tp":3,"q":3,"s":62},{"id":"2117700461751157242","p":23,"t":"Formal verification is the only way, fully convinced. \n\nShoutout to Taelin for seeing it before most of us and building it out.\n\nSomeone needs to do a YouTube series “formal verification for agent acceleration”","ts":1793734109,"l":912,"r":24,"c":37,"v":30745,"tp":2,"q":2,"s":62,"qt":{"id":"2117437607940681849","n":"lauren","h":"poteto","a":"https://pbs.twimg.com/profile_images/2093473719830315008/oo09g1Ov_200x200.jpg","t":"maybe a hot take but i now consider adversarial reviews mostly a waste of tokens. far better to use those tokens on empirical combined with formal verification. i have updated pstack to use the interrogate skill less, and eventually i think i may remove it altogether\n\ni have found very little value with code review in the abstract, agents tend to just get lost in the nitpicks and yagnis","ts":1793671440}},{"id":"2118656836794560632","p":34,"t":"DeepSeek v4.1 Flash 1017 Original weights royal rumble!\n\nEngine (latest version from sources used):\n- mlx-serve by @ddalcu on Apple Silicon \n- TensorFold (@MiaAI_lab recipe) on DGX Spark\n\nMy personal ranking is:\n🥇 DGX Spark\n🥈 2 x M3 Ultra\n🥉 1 x M3 Ultra\n\nOn decode side M3 Ultra wins, but on prefill there is no match. Clearly I bet things change completely adding M5 Ultra in the arena.\n\nSo I think 1 M5 Ultra could be better than 4 x DGX Spark in real usage, but the pro on DGX is CUDA.\n\nEnjoy detailed results and feel free to ask anything, preparing two new benchmarks to push these machines to their very limits!","ts":1793962127,"l":49,"r":8,"c":21,"v":8031,"tp":1,"q":3,"s":60,"ph":[["https://pbs.twimg.com/media/HWb4LWIXoAAn9vT.jpg?name=orig",2400,1350],["https://pbs.twimg.com/media/HWb5UuqWAAAqjW-.jpg?name=orig",3310,1524],["https://pbs.twimg.com/media/HWb5UupXUAAmwom.jpg?name=orig",3286,2216],["https://pbs.twimg.com/media/HWb5UurXMAA-ta8.jpg?name=orig",3272,2166]]},{"id":"2108736734309896625","p":85,"t":"vLLM now supports NVIDIA Vera Rubin. The early results show more than 7.8x the throughput of GB200 on MiniMax M3 on AgentX.\n\n@inferact, @NVIDIAAI, @RedHat_AI, and the vLLM community have been bringing vLLM up on Rubin since it was announced. Here is where things stand.\n\n🧵 1/5","ts":1791596990,"l":463,"r":60,"c":45,"v":84767,"tp":6,"q":3,"s":72,"vd":["https://video.twimg.com/amplify_video/2108736670522859521/vid/avc1/1920x1080/Xn0pblzhSvuQCjkA.mp4?tag=16","https://pbs.twimg.com/amplify_video_thumb/2108736670522859521/img/S0IBFRPTJRwBAWSi.jpg",1920,1080]},{"id":"2118012530215854477","p":39,"t":"AI semiotics: cool paper suggests that AIs are strongly impacted by visual symbols in ways they do not always recognize in their own chains-of-thought. The way they react to those symbols is very different than how a human might interpret them as well. https://arxiv.org/pdf/2610.36449v1","ts":1793808512,"l":421,"r":53,"c":200,"v":30244,"tp":3,"q":3,"s":61,"ph":[["https://pbs.twimg.com/media/HWSxDYWW0AAi8Vk.png?name=orig",862,610],["https://pbs.twimg.com/media/HWSxDYVWkAE6pdQ.png?name=orig",842,763]]},{"id":"2090119165055324518","p":32,"t":"We also released 1-bit Qwen3.8-27B quants which run on 8GB RAM.\n\nThey retain ~77% of accuracy compared to BF16.\n\nWe originally didn't want to release them but were shocked that they worked extremely well on our internal testing!","ts":1787158216,"l":1569,"r":107,"c":32,"v":146936,"tp":1,"q":3,"s":76,"qt":{"id":"2090103470015828184","n":"Unsloth AI","h":"UnslothAI","a":"https://pbs.twimg.com/profile_images/2036313340616777728/OLALWB2__200x200.jpg","t":"We’re releasing new Qwen3.8-27B GGUFs with 10% higher accuracy.\n\nUnsloth Dynamic V3 outperforms others by >10% on Div-300, KLD & more benchmarks.\n\nWe also release 1-bit quants that retain 77% accuracy. Run on 8GB RAM.\n\nBlog: https://unsloth.ai/docs/basics/dynamic-3.0-ggufs\nGGUF: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF","ts":1787154474,"ph":[["https://pbs.twimg.com/media/HQGJu-xaQAAs-Jg.jpg?name=orig",3724,4096]]}},{"id":"2118390190246879554","p":77,"t":"Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.\n\nThe model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work.\n\nGPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.\n\nhttps://openai.com/index/introducing-gpt-6-cyber/","ts":1793898553,"l":21255,"r":1979,"c":1265,"v":3946307,"tp":8,"q":2,"s":60,"vd":["https://video.twimg.com/amplify_video/2118387217588465664/vid/avc1/1920x1080/ggeDE9iY9Trxb8kO.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118387217588465664/img/N2grMUmP0O_EMvS2.jpg",1920,1080]},{"id":"2106029758350032937","p":30,"t":"Decision models in llama.cpp are now available\n\nThe `/v1/systemone` endpoint is available in the latest llama builds. Use it to do Jev-style inference locally, efficiently and privately. Multiple open models are supported with more to come.\n\nhttps://huggingface.co/blog/ggml-org/decision-models-in-llamacpp","ts":1790951597,"l":2256,"r":365,"c":115,"v":116769,"tp":1,"q":3,"s":73},{"id":"2118807846175268873","p":22,"t":"2027 will be the year literally everything will burn down. it will be a glorious bon fire. and once we have paid for the past 40 years of software sins, we will rebuild it. hopefully for the better.\n\nhttps://pwn.ai/blog/firecracker-escape","ts":1793998130,"l":569,"r":36,"c":19,"v":26139,"tp":5,"q":2,"s":59},{"id":"2117299716987789772","p":41,"t":"neither github org level or repo level secrets are actually possible to secure\n\nany person on your team with write access getting hacked means the hacker can exfiltrate those \n\nyou have to use environments with protection rules and put secrets there\n\ncrazy","ts":1793638564,"l":1330,"r":25,"c":117,"v":112751,"tp":5,"q":3,"s":62},{"id":"2110875974296654062","p":15,"t":"This is interesting work, but I don't think it's a significant update to the big picture of interpretability.\n\nI'm on board with the broad idea of doing exploratory interpretability with neurons, e.g. neuron-based attribution graphs, for pragmatic problems like why did a model do a sketchy thing? I think this work provides some reasonable evidence, in addition to good earlier work from\nTransluce, that this might actually work for practical purposes, and has some advantages over the standard alternatives like SAEs/transcoders. I do not agree with some people's grander inferences like that this is substantial progress towards ambitious reverse engineering, solving interpretability etc.\n\nI expect that a neuron-based attribution graph will still have significant error and be missing a lot just like transcoders, once you prune them down to the most important components, even if they solve error terms. Which is fine for pragmatic purposes, and may suffice for getting the gist of what's going on, but I remain pessimistic about really fully understanding things and being confident there's nothing we're missing. I do think that using neurons is just vastly easier to, for example, do on a frontier model, which is great. I'm also excited to see more innovation in using agents to interpret neurons. I feel like this is an area the field has somewhat neglected, and where there's probably a very high ceiling on how far things can go. The SwiGLU double negative thing is also a useful point.\n\nI still think superposition is real and this work doesn’t really affect my cruxes. Most importantly, it seems highly likely to be true that the model can represent more concepts in the residual stream than it has dimensions, ie representational superposition, which I think the authors agree with me on. For neuron/computational superposition, it seems to me like language models have far more facts and memorized things in them than they have neurons. And that superposition is an efficient way to pack in more facts than you have neurons by representing them as linear combinations of neurons, especially for arbitrary unstructured values. Facts/memorized content is largely an unstructured database of key-value pairs IMO and I don’t see how this is compatible with a monosemantic view of neurons. I expect that weak neuron activations contain lots of semantic superposition-y info, which, if ablated, would meaningfully damage network performance. As I understand the work, it does not properly address what's going on with weak neuron activations in LLMs (I am skeptical of how much vision results will transfer, eg I expect small image models memorise far less). This picture is compatible with the highest activations of a neuron being mono semantic (though IMO the post defines monosemantic too loosely, I think I’d call a fair few of those polysemantic). It’s also plausible that the most important concepts are likely to correspond to a neuron (or a handful), which would be great, and may suffice for practical purposes!\n\nSimilarly, my impression is that SAEs generally don't find things that are a\nsingle neuron, or even very sparse in the neuron basis, and are often a better basis for certain tasks than the neuron basis, though the evidence base is kinda weak here since we largely focus on residual stream or layer output SAEs, I’d love to see work on this in Goodfire’s forthcoming paper! I also think residual stream SAEs are extremely useful, as they give a good bottleneck of the model’s information at this token and layer, and I don’t see a good way to replicate them in the neuron framework.\n\nAnd kudos to Nick for updating a bunch of his historical vision work! Revisiting old results with modern tools and more rigour seems great.","ts":1792107025,"l":1763,"r":175,"c":39,"v":136426,"tp":3,"q":3,"s":70,"qt":{"id":"2110757819427574056","n":"Nick","h":"nickcammarata","a":"https://pbs.twimg.com/profile_images/1753264923365523456/mUCvwn7v_200x200.jpg","t":"interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had\n\nwith modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works\n\nhttps://already-clean.goodfire.com/","ts":1792078854}},{"id":"2067301278976180531","p":83,"t":"For WWDC, we worked with Apple to run Kimi K2.6, a 1T-parameter model, across a cluster of four Mac Studios using a preview version of LM Studio.\n\nWe showcased secure remote access from a MacBook Neo and iPhone using LM Link.\n\nA glimpse of your own private, frontier-scale AI.","ts":1781718008,"l":4413,"r":314,"c":130,"v":406685,"tp":1,"q":3,"s":76,"ph":[["https://pbs.twimg.com/media/HLCFmjzXUAAj38o.jpg?name=orig",2048,1536],["https://pbs.twimg.com/media/HLCF0KzXsAA0Cmk.jpg?name=orig",3024,4032],["https://pbs.twimg.com/media/HLCF1cqWoAAbTyz.jpg?name=orig",3024,4032]]},{"id":"2064202168618422396","p":31,"t":"It's very cool that Apple shipped a 20B parameter on-device. \n\nYou can't put 20B parameters in RAM at any reasonable precision. To make it work they are using pretty exotic architecture by today's standards.\n\nA small model predicts from the query (or prompt) which experts to load from Nand into RAM. The key distinction from a typical MoE is that you do this once per query and then generate all the tokens with the same experts (instead of switching the experts for every token).","ts":1780979122,"l":3186,"r":290,"c":73,"v":230941,"tp":1,"q":3,"s":76,"ph":[["https://pbs.twimg.com/media/HKWEx5Hb0AARN-b.jpg?name=orig",1926,1104]]},{"id":"1995452641430651132","p":81,"t":"🚀 Launching DeepSeek-V3.2 & DeepSeek-V3.2-Speciale — Reasoning-first models built for agents!\n\n🔹 DeepSeek-V3.2: Official successor to V3.2-Exp. Now live on App, Web & API.\n🔹 DeepSeek-V3.2-Speciale: Pushing the boundaries of reasoning capabilities. API-only for now.\n\n📄 Tech report: https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/main/assets/paper.pdf\n\n1/n","ts":1764587958,"l":15892,"r":2388,"c":940,"v":5336652,"tp":8,"q":3,"s":76,"ph":[["https://pbs.twimg.com/media/G7FDq18aMAAWWvh.jpg?name=orig",4096,2286]]},{"id":"1958417062008918312","p":81,"t":"Introducing DeepSeek-V3.1: our first step toward the agent era! 🚀\n\n🧠 Hybrid inference: Think & Non-Think — one model, two modes\n⚡️ Faster thinking: DeepSeek-V3.1-Think reaches answers in less time vs. DeepSeek-R1-0528\n🛠️ Stronger agent skills: Post-training boosts tool use and multi-step agent tasks\n\nTry it now — toggle Think/Non-Think via the \"DeepThink\" button: https://chat.deepseek.com/\n\n1/5","ts":1755757987,"l":14601,"r":1733,"c":501,"v":2100466,"tp":8,"q":3,"s":76},{"id":"1906692961343009163","p":3,"t":"It's been a while, so you might be yearning...\n...for a StatQuest video about Reinforcement Learning!\nBAM!\nhttps://youtu.be/Z-T0iJEXiwM","ts":1743426001,"l":425,"r":38,"c":14,"v":32135,"tp":0,"q":3,"s":76},{"id":"1891745487071609327","p":81,"t":"🚀 Introducing NSA: A Hardware-Aligned and Natively Trainable Sparse Attention mechanism for ultra-fast long-context training & inference!\n\nCore components of NSA:\n• Dynamic hierarchical sparse strategy\n• Coarse-grained token compression\n• Fine-grained token selection\n\n💡 With optimized design for modern hardware, NSA speeds up inference while reducing pre-training costs—without compromising performance. It matches or outperforms Full Attention models on general benchmarks, long-context tasks, and instruction-based reasoning.\n\n📖 For more details, check out our paper here: https://arxiv.org/abs/2502.11089","ts":1739862245,"l":15085,"r":2042,"c":876,"v":2616782,"tp":3,"q":3,"s":76,"ph":[["https://pbs.twimg.com/media/GkDUL57aAAARneI.png?name=orig",2268,1208],["https://pbs.twimg.com/media/GkDUMe3aAAMyB04.png?name=orig",1612,928],["https://pbs.twimg.com/media/GkDUNRvbsAAoUbx.png?name=orig",2168,1038],["https://pbs.twimg.com/media/GkDUOFmaAAA8b8u.png?name=orig",2166,1158]]},{"id":"1889784295213998406","p":3,"t":"I'm still wrapping my brain around the fact that this happened. Will someone pinch me? Or, better, just sign up for this free course https://bit.ly/4hDKXYg and tell me if it's real.","ts":1739394661,"l":410,"r":31,"c":20,"v":20427,"tp":0,"q":3,"s":76,"qt":{"id":"1889766176059994166","n":"Andrew Ng","h":"AndrewYNg","a":"https://pbs.twimg.com/profile_images/733174243714682880/oyG30NEH_200x200.jpg","t":"New short course: Attention in Transformers: Concepts and Code in PyTorch.\n\nLast week we released a course on how LLM transformers work. This week, go deeper and learn about the technical ideas behind the attention mechanism, and see how to code it in PyTorch. This course is built with @joshuastarmer, Founder and CEO of StatQuest.\n\nThe attention mechanism was a breakthrough that led to transformers, the architecture powering large language models like ChatGPT. Transformers, introduced in the 2017 paper: \"Attention is All You Need\" by Viswani and others, took off because of its highly scalable design. \n\nIn this course, you’ll learn how the attention mechanism, a key element of transformer-based LLMs, works and implement it in PyTorch. You'll develop deep intuition about building reliable, functional, and scalable AI applications.\n\nWhat you will do:\n- Understand the evolution of the attention mechanism, a key breakthrough that led to transformers.\n- Learn the relationships between word embeddings, positional embeddings, and attention.\n- Learn about the Query, Key, and Value matrices, and how to produce and use them in attention.\n- Walk through the math required to calculate self-attention and masked self-attention to learn why and how they work.\n- Understand the difference between self-attention and masked self-attention and how one is used in the encoder to build context-aware embeddings and the other is used in the decoder for generative outputs.\n- Learn the details of the encoder-decoder architecture, cross-attention, and multi-head attention and how they are all incorporated into a transformer.\n- Use PyTorch to code a class that implements self-attention, masked self-attention, and multi-head attention.\n\nThere're lots of exciting technical details in this course.  Please sign up here: https://www.deeplearning.ai/short-courses/attention-in-transformers-concepts-and-code-in-pytorch","ts":1739390341,"vd":["https://video.twimg.com/ext_tw_video/1889765173700993026/pu/vid/avc1/1280x720/OytWYxHEYM8dpsFF.mp4?tag=12","https://pbs.twimg.com/ext_tw_video_thumb/1889765173700993026/pu/img/vdCdZxJAcO9tE83w.jpg",1920,1080]}},{"id":"1777320491222712762","p":3,"t":"Following me sometimes has its perks...Here's the matrix math behind transformer neural networks!!! BAM! https://youtu.be/KphmOJnLAdI","ts":1712581200,"l":433,"r":48,"c":4,"v":37241,"tp":0,"q":3,"s":76},{"id":"2116662918091858099","p":52,"t":"Feels like it's been some time since I've learned something new about non-determinism on GPUs 😆 Pretty cool!","ts":1793486740,"l":403,"r":12,"c":9,"v":22843,"tp":6,"q":3,"s":63,"qt":{"id":"2116241086050115692","n":"div (fp32)","h":"divagr1925","a":"https://pbs.twimg.com/profile_images/2103115932990332929/wKj91D1E_200x200.jpg","t":"1/4\n\nNew writeup: you can fingerprint the exact GPU you're running on with a default torch.index_add_ call, even among identical GPUs.\n\nThis works because floating-point maths depends on order, and every GPU has its own. This fingerprint survives reboots and can't be faked.","ts":1793386167,"vd":["https://video.twimg.com/amplify_video/2116240958836948992/vid/avc1/1280x720/jOU72hsx-EbqzL4n.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2116240958836948992/img/_OdQZYS10DmhsZkL.jpg",1280,720]}},{"id":"2107502286758895878","p":78,"t":"Meet EmbeddingGemma 2, our first natively multimodal open model for on-device embeddings.\n\nIt expands beyond text to unify code, images, audio, and video in a shared space. 🧵","ts":1791302675,"l":4491,"r":537,"c":292,"v":1874053,"tp":1,"q":3,"s":72,"vd":["https://video.twimg.com/tweet_video/HT9Uhe6WYAA8Sg9.mp4","https://pbs.twimg.com/tweet_video_thumb/HT9Uhe6WYAA8Sg9.jpg",960,540]},{"id":"2116599971176935519","p":46,"t":"I like to think that i know reasonably well what to expect from floats on gpus, but the consistency in this one still surprised me!","ts":1793471732,"l":326,"r":16,"c":13,"v":25701,"tp":6,"q":3,"s":63,"qt":{"id":"2116241086050115692","n":"div (fp32)","h":"divagr1925","a":"https://pbs.twimg.com/profile_images/2103115932990332929/wKj91D1E_200x200.jpg","t":"1/4\n\nNew writeup: you can fingerprint the exact GPU you're running on with a default torch.index_add_ call, even among identical GPUs.\n\nThis works because floating-point maths depends on order, and every GPU has its own. This fingerprint survives reboots and can't be faked.","ts":1793386167,"vd":["https://video.twimg.com/amplify_video/2116240958836948992/vid/avc1/1280x720/jOU72hsx-EbqzL4n.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2116240958836948992/img/_OdQZYS10DmhsZkL.jpg",1280,720]}},{"id":"2118634626679775549","p":21,"t":"This is a conspiracy theory but I'm like 95% sure I'm right. Reasoning traces for Opus 5.5 read like a shitty quantized open weight model then SOMEHOW the right answer comes out in the end.","ts":1793956831,"l":4962,"r":80,"c":213,"v":419892,"tp":3,"q":2,"s":59,"qt":{"id":"2118616574676570529","n":"Theo - t3.gg","h":"theo","a":"https://pbs.twimg.com/profile_images/2106484435382333440/vXOVE8EX_200x200.jpg","t":"@RhysSullivan Ignore reasoning traces from Opus 5.5. They do not represent what the model is actually thinking. I swear they're just there to fuck up distillation","ts":1793952528}},{"id":"2113654775665631551","p":14,"t":"How a modern transistor works:","ts":1792769542,"l":2902,"r":201,"c":36,"v":150834,"tp":6,"q":3,"s":67,"vd":["https://video.twimg.com/amplify_video/2113653955956023296/vid/avc1/3840x2160/oHFZW_WoD9OyrqUf.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2113653955956023296/img/OUofKuEY19hyWH75.jpg",3840,2160]},{"id":"2112592787338408445","p":54,"t":"This is a very big deal. DeepSeeks inference system is on par with OpenAIs stack and better what Anthropic is running (at least for their user size).","ts":1792516345,"l":783,"r":59,"c":21,"v":80391,"tp":6,"q":3,"s":68,"qt":{"id":"2112476496283775135","n":"DeepSeek","h":"deepseek_ai","a":"https://pbs.twimg.com/profile_images/1717417613775757312/Uk1zNOj4_200x200.jpg","t":"🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!\n\nPowering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.\n\n🔗 http://github.com/deepseek-ai/DeepInfer\n\n1/4","ts":1792488619}},{"id":"2116957135661945283","p":0,"t":"Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:\nhttps://hackertube.org/\n\n- Excellent recs without any logins (state is kept in local browser session)\n- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads\n- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version\n- HackerTube Longs ✊\n\nA bit more on the process:\nI use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only \"manual work\" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.\n\nMany awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.","ts":1793556886,"l":51761,"r":4759,"c":2332,"v":5263986,"tp":2,"q":2,"s":62,"ph":[["https://pbs.twimg.com/media/HWDnuuCaIAA__wa.jpg?name=orig",2482,1442]]},{"id":"2117495577248485856","p":13,"t":"👀","ts":1793685261,"l":30,"r":1,"c":5,"v":2734,"tp":3,"q":3,"s":61,"qt":{"id":"2117313742891327994","n":"Rikiya Takehi","h":"rikiyatakehi","a":"https://pbs.twimg.com/profile_images/1731888727897567232/97PMSyds_200x200.jpg","t":"We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.\nIt directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.\n\nWe call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵\nBlog: https://rikiyat.github.io/blog/weight-space-programming/","ts":1793641908,"ph":[["https://pbs.twimg.com/media/HWI1oKUWUAAWoZp.jpg?name=orig",1200,625]]}},{"id":"2118894224733249655","p":21,"t":"I gave Adam access to the T3 Code anonymized telemetry data to make more posts like this :)\n\nAnyways, more reasons to never use \"max\" reasoning:","ts":1794018724,"l":872,"r":9,"c":53,"v":68531,"tp":2,"q":3,"s":58,"qt":{"id":"2118871687848477030","n":"Adam Holter","h":"AdamHoltererer","a":"https://pbs.twimg.com/profile_images/2093828048684544001/5JR0MimQ_200x200.jpg","t":"Opus 5.5 fails twice as many T3 turns on max as on high\n\nMy first guess would be that it's because turns on max, take longer, but no: max fails more in every turn-length bucket, and most of the extra failures happen in the first 30 seconds.\n\nSol shows the same pattern, smaller (1.4x)\n\nSo why is this? ¯\\_(ツ)_/¯","ts":1794013351,"vd":["https://video.twimg.com/amplify_video/2118871322369359872/vid/avc1/1080x1080/48FTlfR_Bey2ViOj.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118871322369359872/img/7JureiSI9f7NTXz9.jpg",1080,1080]}},{"id":"2118400768957202439","p":17,"t":"we're recording an AAv5 pod with @_micah_h and @grmcameron today and this reminds me why you basically need @ArtificialAnlys AA-* rebuilds of every single eval \n\nbecause you -CANNOT- trust labs not to intentionally or accidentally benchmaxx, you need a few trusted third parties to basically redo the entire eval in spirit but not in letter, to have a quick sniff test that model claims hold up under minor perturbations (eg starting from \"randomizing order in multiple choice questions\" to coming up with completely new problems that superficially look different but actually test the same underlying logic and reasoning steps). not enough people are reading bench traces like this 👇  and the systematic way to catch this benchmaxxing is basically to have altbenches for everything","ts":1793901075,"l":66,"r":0,"c":102,"v":7282,"tp":3,"q":3,"s":59,"qt":{"id":"2117613618065793160","n":"Joachim Schaeffer","h":"JSchaeff3r","a":"https://pbs.twimg.com/profile_images/2020470608778436609/GaAB7M2l_200x200.jpg","t":"Kimi K3.1 recalls instructions it was never given on Terminal Bench. They match the hidden tests. K3.1 while solving: \"Instructions likely distilled pitfalls from previous agents.\" 🤯\n\nWhen we edit a task to contradict them, K3.1 refuses the new instructions and fails. Other models follow the edit. \n\nIs it terminal(ly) benchmaxxed? 🧵","ts":1793713404,"ph":[["https://pbs.twimg.com/media/HWM_uhsXIAAz1sN.jpg?name=orig",1100,1002]]}},{"id":"2114857098357911617","p":46,"t":"can't believe it's actually working","ts":1793056199,"l":2100,"r":102,"c":22,"v":68402,"tp":3,"q":2,"s":65,"ph":[["https://pbs.twimg.com/media/HVl7PdrXoAA-x8T.png?name=orig",751,479]],"qt":{"id":"2114836071028785289","n":"Andrew Curran","h":"AndrewCurran_","a":"https://pbs.twimg.com/profile_images/1596945208058744833/_X3LT7fb_200x200.jpg","t":"OpenAI has announced to three new discoveries in mathematics, computer science, and physics:\n- Local smoothing and the Kakeya conjecture\n- The Exponential Time Hypothesis is false\n- Existence of quantum chromodynamics and completions of the Standard Model","ts":1793051185,"ph":[["https://pbs.twimg.com/media/HVloMu2aYAAYX5k.jpg?name=orig",1080,753]]}},{"id":"2118830837840183628","p":21,"t":"New prompt suffix I’ve been liking lately:\n\n“Be thorough in your analysis and bold in your recommendations”","ts":1794003612,"l":3881,"r":76,"c":116,"v":121553,"tp":2,"q":2,"s":58},{"id":"2116162363313525191","p":32,"t":"We're releasing new Unsloth v3.1 GGUF quants for DeepSeek-V4.1-Flash-1017. We ran benchmarks and v3 + v3.1 do better than GSQ-RCO.\n\nWe also cannot reproduce numbers ran by GSQ-RCO - AIME25 for 1bit is not 100% but rather 91.88%. Livecodebench and GPQA are similarly incorrect.\n\nThere are also claims that KLD is not a good metric, and that people can \"game\" KLD. The pirmary goal of quantization is to minimize KLD but also minimize disk space. If you get a KLD of 0, this means your quantized model == 16bit model, as logits are perfectly modelled.\n\nWe ran benchmarks on multiple benchmarks including AIME25, Aider Polyglot, LiveCodeBench v6, HMMT 2026, GPQA Diamond, MMLU-Pro and IFEval, and also recorded reasoning length to see if lower bit quants think for longer.\n\nIn https://unsloth.ai/docs/models/deepseek-v4.1-flash#quantization-increases-reasoning-lengths, we show that lower bit quants do think somewhat longer than the unquantized model - we plan to investigate and make quants even better.\n\nHowever, we're still much ahead vs GSQ-RCO on even reasoning length as well.\n\nWe provide full benchmarking data, benchmarking methods (for eg we ran AIME25 16 times per question and took an average) and more in our updated blog at https://unsloth.ai/docs/models/deepseek-v4.1-flash","ts":1793367398,"l":182,"r":18,"c":30,"v":14443,"tp":1,"q":3,"s":63,"ph":[["https://pbs.twimg.com/media/HV4b5AOagAAkZb2.jpg?name=orig",2964,858]],"qt":{"id":"2116156587404562491","n":"Unsloth AI","h":"UnslothAI","a":"https://pbs.twimg.com/profile_images/2036313340616777728/OLALWB2__200x200.jpg","t":"We’re releasing new DeepSeek-V4.1-Flash-1017 GGUFs, ranking highest across benchmarks.\n\nDynamic V3.1 leads across nearly all benchmarks: GPQA Diamond, AIME25, LiveCodeBench v6, MMLU-Pro, Aider & more.\n\nRun on 64GB+ RAM.\n\nGGUF: https://huggingface.co/unsloth/DeepSeek-V4.1-Flash-1017-GGUF\nBlog: https://unsloth.ai/docs/models/deepseek-v4.1-flash","ts":1793366021,"ph":[["https://pbs.twimg.com/media/HV4XBGya0AA3qin.jpg?name=orig",3724,4096]]}},{"id":"2117316068226028025","p":44,"t":"you thought LLMs can only do text optimization and discovery for prompts, harnesses, and code?\n\nturns out they’re starting to be good at natively reasoning about and writing out neural network weights directly too\n\nin the limit, where does this leave symbolic local optimization heuristics like gradient descent…??","ts":1793642463,"l":535,"r":47,"c":49,"v":47906,"tp":3,"q":2,"s":61,"qt":{"id":"2117313742891327994","n":"Rikiya Takehi","h":"rikiyatakehi","a":"https://pbs.twimg.com/profile_images/1731888727897567232/97PMSyds_200x200.jpg","t":"We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.\nIt directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.\n\nWe call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵\nBlog: https://rikiyat.github.io/blog/weight-space-programming/","ts":1793641908,"ph":[["https://pbs.twimg.com/media/HWI1oKUWUAAWoZp.jpg?name=orig",1200,625]]}},{"id":"2116735229792522518","p":41,"t":"this is a big issue with the \"bash for everything\" approach\n\nif you look you'll see it using python for edits then && run tests && format for every single change\n\nit's too loose","ts":1793503980,"l":1526,"r":19,"c":100,"v":120104,"tp":2,"q":3,"s":62,"qt":{"id":"2116655759278973340","n":"Tejas Bhakta","h":"tejasybhakta","a":"https://pbs.twimg.com/profile_images/1983331868809539584/nXr2234a_200x200.jpg","t":"models have no sense of time. they'll trigger a 30 min rebuild for a 2 line change instead of thinking ahead and batching the work\n\ncurrent frontier RL rewards token efficiency\nfuture RL will reward time efficiency\n\nthe fast takeoff will feel absurd after this","ts":1793485033}},{"id":"2070237256070389897","p":48,"t":"A super long overdue (3+ years?) post on scaling laws.\n\nCompute is expensive. Scaling laws are a way to help us reason about the optimal compute allocation between data and model size before committing to a large run. \n\nThe post covers what scaling laws predict, how compute-optimal allocation works, why Kaplan et al. and Chinchilla disagree, and how data limits + fitting details make extrapolation tricky.\n\nhttps://lilianweng.github.io/posts/2026-06-24-scaling-laws/","ts":1782417999,"l":4679,"r":591,"c":73,"v":486445,"tp":0,"q":3,"s":75},{"id":"2039752638384709661","p":30,"t":"Let me demonstrate the true power of llama.cpp:\n\n- Running on Mac Studio M2 Ultra (3 years old)\n- Gemma 4 26B A4B Q8_0 (full quality)\n- Built-in WebUI (ships with llama.cpp)\n- MCP support out of the box (web-search, HF, github, etc.)\n- Prompt speculative decoding\n\nThe result: 300t/s\n\n(realtime video)","ts":1775149900,"l":3289,"r":258,"c":130,"v":786275,"tp":1,"q":3,"s":75,"vd":["https://video.twimg.com/amplify_video/2039751265962352640/vid/avc1/2160x2368/mZ3R_upoxLK62ItC.mp4?tag=21","https://pbs.twimg.com/amplify_video_thumb/2039751265962352640/img/txbzNCUJoJFWxeL6.jpg",2696,2956]},{"id":"2038632534414680223","p":30,"t":"llama.cpp at 100k stars\n\nnow that 90% of the code worldwide is being written by AI agents, I predict that within 3-6 months, 90% of all AI agents will be running locally with llama.cpp 😄\n\nJokes aside, I am going to use this small milestone as an opportunity to reflect a bit on the project and the state of AI from the perspective of local applications. There is a lot to say and discuss and yet it feels less and less important to try to make a point. Opinions about viability of local LLMs are strongly polarized, details are overlooked, the scientific approach is lacking. Arguments are predominantly based on vibes and hype waves.\n\nOne thing is clear though - local LLMs are used more and more. I expect this trend to continue and likely 2026 will end up being one of the most important years for the local AI movement.\n\nI admit that I didn't expect the agentic era to come so quickly to the local LLM space. One year ago, the available models were too computationally expensive for doing long-context tasks. There wasn't an obvious path towards meaningful agentic applications. The memory and compute requirements were huge. Last summer, with the release of gpt-oss, things started to change. It was the first time we saw a glimpse of tool calling that actually works well within the resource constraints of our daily devices. Later in the year, even better models were released and by now, useful local agentic workflows are a reality.\n\nComparing local vs hosted capabilities at a given moment of time is pointless. To try put things into perspective:\n\n- We don't need frontier intelligence to automate searches and sending emails\n- We don't need trillion parameter models to be able to summarize articles or technical documents\n- We don't need massive GPU data centers to control our home appliances or turn the lights off in the garage\n\nI believe that there is a certain level of intelligence we as humans can comprehend and meaningfully utilize to improve our working process. Beyond that level, access to more intelligence becomes unnecessary at best and counterproductive at worst. I also believe that that level of useful artificial intelligence is completely within reach locally and it has always been just a matter of implementing the right software stack to bring it to the end user.\n\nWith llama.cpp, I am confident that we continue to be on the right track of building that software stack!\n\nThe llama.cpp project is going stronger than ever. With more than 1500 contributors, the project keeps growing steadily.\n\nFrom technical point of view, I think that llama.cpp + ggml is the only solution that actually makes sense. That is, the software stack must run efficiently on every possible device, hardware and operating system. The technology is too important to be vendor-locked. It has to be developed in the open, by the community, together with the independent hardware vendors. This is the only right way to build something that will truly make a difference in the long run.\n\nI won't try to convince you about what is currently and will be possible with local AI. We will just continue to build as usual. I am confident that after the smoke clears and we look objectively at what we have built together, the benefits will be obvious to everyone.\n\nBig shoutout to all llama.cpp maintainers. I feel extremely lucky to be able to work together with so many talented contributors. Every day I learn something new and I feel there is so much more cool stuff that we are going to build. Also, I am really thankful that the project continues to have reliable partners to support it!\n\nCheers!","ts":1774882846,"l":3288,"r":398,"c":295,"v":479698,"tp":1,"q":3,"s":75,"ph":[["https://pbs.twimg.com/media/HEqpwjaWYAEQQQQ.jpg?name=orig",2030,990],["https://pbs.twimg.com/media/HEqp851XoAAw5n0.jpg?name=orig",2038,1654]]},{"id":"1965828670167331010","p":52,"t":"Apologies that I haven't written anything since joining Thinking Machines but I hope this blog post on a topic very near and dear to my heart (reproducible floating point numerics in LLM inference) will make up for it!","ts":1757525052,"l":2867,"r":174,"c":88,"v":575475,"tp":6,"q":3,"s":75,"qt":{"id":"1965826369721623001","n":"Thinking Machines","h":"thinkymachines","a":"https://pbs.twimg.com/profile_images/2075275500420001792/Q9J4M1DL_200x200.jpg","t":"Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference”\n\nWe believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to prompt engineering. Here we share what we are working on and connect with the research community frequently and openly.\n\nThe name Connectionism is a throwback to an earlier era of AI; it was the name of the subfield in the 1980s that studied neural networks and their similarity to biological brains.\n\nhttps://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/","ts":1757524504,"ph":[["https://pbs.twimg.com/media/G0gES2uaEAEc01p.jpg?name=orig",1788,939]]}},{"id":"1881353126210687089","p":47,"t":"We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive - truly open, frontier research that empowers all. It makes no sense. The most entertaining outcome is the most likely.\n\nDeepSeek-R1 not only open-sources a barrage of models but also spills all the training secrets. They are perhaps the first OSS project that shows major, sustained growth of an RL flywheel.\n\nImpact can be done by \"ASI achieved internally\" or mythical names like \"Project Strawberry\". \nImpact can also be done by simply dumping the raw algorithms and matplotlib learning curves.\n\nI'm reading the paper:\n\n> Purely driven by RL, no SFT at all (\"cold start\"). Reminiscent of AlphaZero - master Go, Shogi, and Chess from scratch, without imitating human grandmaster moves first. This is the most significant takeaway from the paper.\n> Use groundtruth rewards computed by hardcoded rules. Avoid any learned reward models that RL can easily hack against.\n> Thinking time of the model steadily increases as training proceeds - this is not pre-programmed, but an emergent property! \n> Emergence of self-reflection and exploration behaviors.\n> GRPO instead of PPO: it removes the critic net from PPO and uses the average reward of multiple samples instead. Simple method to reduce memory use. Note that GRPO was also invented by DeepSeek in Feb 2024 ... what a cracked team.","ts":1737384513,"l":8490,"r":1461,"c":215,"v":1403292,"tp":3,"q":3,"s":75,"ph":[["https://pbs.twimg.com/media/GhvhiJsbsAAm5pq.jpg?name=orig",2000,1648]]},{"id":"1872242657348710721","p":81,"t":"🚀 Introducing DeepSeek-V3!\n\nBiggest leap forward yet:\n⚡ 60 tokens/second (3x faster than V2!)\n💪 Enhanced capabilities\n🛠 API compatibility intact\n🌍 Fully open-source models & papers\n\n🐋 1/n","ts":1735212408,"l":12891,"r":2084,"c":646,"v":7436304,"tp":8,"q":3,"s":75,"ph":[["https://pbs.twimg.com/media/GfuKUsuagAEfvM2.jpg?name=orig",1280,1091]],"vd":["https://video.twimg.com/tweet_video/GfuKT4BbUAA3Wxh.mp4","https://pbs.twimg.com/tweet_video_thumb/GfuKT4BbUAA3Wxh.jpg",1132,682]},{"id":"2117767095090053475","p":85,"t":"vLLM v0.33.0 is out! 🚀\n\n927 commits · 389 contributors · 122 first-timers\n\n🤖 New models: Reflection Beam-1.0, Step-5 Preview, Nemotron Parse 3.5, MiniCPM-OCR 1.0\n🔥 PyTorch 2.14 — CUDA 13.0-only wheels, libtorch ABI-stable\n🎲 Spec decoding: XPress refiner, FLy lossy verify (1.04–1.31x), DSpark 1373→3390 tok/s on GB300\n⚡ NVIDIA decode: 1.5–2x latency decode on GB300, FA4 sparse MLA, 39–48% lower TTFT\n🚀 New AWS EFA image + shorter vllm/vllm name\n🐧 ROCm 10.0 default, MonoKernel 2.32x P50\n\nFull notes ➡️ http://github.com/vllm-project/vllm/releases/tag/v0.33.0","ts":1793749996,"l":287,"r":29,"c":21,"v":18443,"tp":6,"q":2,"s":60,"ph":[["https://pbs.twimg.com/media/HWPR8h-aIAAUbBp.jpg?name=orig",1200,675]]},{"id":"2118707871139696911","p":84,"t":"Qwen3.8-Flash-Next GGUFs now run faster with less thinking and less token use.\n\nUnsloth Dynamic V3.1 GGUFs bring thinking & token use to match 95% of full precision Qwen3.8-Flash.\n\nResults: 50% fewer tokens + 15% higher accuracy vs same size quants.\n\nGGUF: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF","ts":1793974294,"l":1126,"r":93,"c":60,"v":75603,"tp":1,"q":3,"s":58,"ph":[["https://pbs.twimg.com/media/HWcpVvua8AATCUx.jpg?name=orig",3724,4096]],"qt":{"id":"2118033433045991660","n":"Unsloth AI","h":"UnslothAI","a":"https://pbs.twimg.com/profile_images/2036313340616777728/OLALWB2__200x200.jpg","t":"We’re releasing new Qwen3.8-Flash-Next GGUFs with 15% higher accuracy.\n\nUnsloth Dynamic V3.1 outperforms others by >15% on Terminal Bench, DeepSWE & more benchmarks (256K context).\n\nRun on 32GB+ RAM setups.\n\nGGUF: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF\nBlog: https://unsloth.ai/docs/models/qwen3.8-next","ts":1793813496,"ph":[["https://pbs.twimg.com/media/HWTEDzGboAAl8mB.jpg?name=orig",3724,4096]]}},{"id":"2112907856018309586","p":77,"t":"OpenAI models helped researchers—operating as part of our Patch the Planet initiative—discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites.\n\nOver the past two months, we led a coalition of maintainers, researchers, and security providers to prepare fixes and protections ahead of disclosure. This duration was necessary because the fix was quite intricate and required a major change to the PHP engine. Every site using PHP can now safely upgrade their version to stay secure, in addition to applying mitigations from various security vendors.\n\nDue to the severity of these flaws and PHP’s widespread use, we consider this bug to be one of the most impactful vulnerabilities in recent memory. It also marks a new phase, where AI can find security bugs with internet-scale impact. \n\nThis is an example of the Defender’s Window in action: to our knowledge, frontier AI was able to find this vulnerability ahead of attackers, giving defenders time to prepare.\nhttps://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities/","ts":1792591463,"l":14442,"r":2866,"c":626,"v":5988864,"tp":5,"q":3,"s":67},{"id":"2119168170045124855","p":34,"t":"Look at herdr managing multiple omp sessions connected to DeepSeek v4.1 Flash running on 4xDGX Spark 🤯","ts":1794084038,"l":12,"r":1,"c":1,"v":945,"tp":1,"q":2,"s":43,"vd":["https://video.twimg.com/amplify_video/2119168015086297088/vid/avc1/3790x2160/nBrTEZngtm8RqT3l.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2119168015086297088/img/Lb62RgFl445IbvFa.jpg",4096,2334]},{"id":"2119161981303398508","p":23,"t":"Permanent underclass is dead \n\nOpenAI shaking","ts":1794082563,"l":84,"r":1,"c":9,"v":4279,"tp":8,"q":2,"s":48,"qt":{"id":"2118489527362912566","n":"Thinking Machines","h":"thinkymachines","a":"https://pbs.twimg.com/profile_images/2075275500420001792/Q9J4M1DL_200x200.jpg","t":"Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.\n\nInkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.\n\nhttps://thinkingmachines.ai/news/inkling-2-small","ts":1793922237}},{"id":"2119157991048126769","p":21,"t":"Benchmarks are an art form","ts":1794081611,"l":358,"r":8,"c":33,"v":19586,"tp":2,"q":2,"s":43,"qt":{"id":"2118880329490637049","n":"Justin Lo","h":"loresjpg","a":"https://pbs.twimg.com/profile_images/2117764091595075584/gso3afey_200x200.jpg","t":"Nice DeepSWE score! Does it survive a lowercase filename?\n\nIntroducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.\n\nModels clustered around 74% on DeepSWE end up 23 pp apart.\n\nGood at coding, or good at DeepSWE? 🧵","ts":1794015412,"ph":[["https://pbs.twimg.com/media/HWfGLoKaUAAi2sL.jpg?name=orig",2048,1152]]}},{"id":"2119141295071367211","p":39,"t":"Now there might be things to verify beyond errors - points of view, opinions, approaches, etc. But the idea that the job of humans should be checking AI work for mistakes is likely not true today, and good riddance.","ts":1794077631,"l":55,"r":2,"c":6,"v":6006,"tp":2,"q":2,"s":39},{"id":"2119113676968063221","p":34,"t":"1M context benchmark on 4 x DGX Sparks!\nLook at these numbers! 🔥\n\nGreat job @christopherowen 🙌\n\nBut there is a new contender: @hunterallen let's test and compare!","ts":1794071046,"l":40,"r":6,"c":11,"v":2911,"tp":1,"q":2,"s":50,"ph":[["https://pbs.twimg.com/media/HWiaMGjXEAAkmDU.jpg?name=orig",2400,1350]]},{"id":"2119100152241877118","p":23,"t":"A hand-drawn valley on plain paper, with origami cranes landing on Inkling. Generated with Inkling 2 Small in Blender; revised with improvements to the water and chimney.\n\nI love these demos so much","ts":1794067821,"l":48,"r":2,"c":9,"v":4213,"tp":8,"q":2,"s":45,"vd":["https://video.twimg.com/amplify_video/2119099964475465728/vid/avc1/1920x1080/DecBnwG6VpBk-6jF.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2119099964475465728/img/7NCU07wkAUrrcasS.jpg",1920,1080]},{"id":"2119091307503079739","p":34,"t":"Here DSv4.1 Flash recipe by @christopherowen in action!\n\nDeepSeek Harness with a basic prompt: \"Use Blender MCP to create animation of a lighthouse in the middle of a storm.\" 🤯\n\nAvg speed: 73 t/s\nTime:  49 mins\nTokens: ~22M - 300K output\nCache Hit: 99%\nContext used: 295K\n\nGive it a try!","ts":1794065713,"l":79,"r":3,"c":15,"v":4893,"tp":1,"q":2,"s":54,"vd":["https://video.twimg.com/amplify_video/2119090845496287232/vid/avc1/1280x720/GtJ0bQnR4bPofP-U.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2119090845496287232/img/B7YEHE50BxGKYMQ-.jpg",1280,720]},{"id":"2119084530883801490","p":39,"t":"AI could do 1000x more research but the jagged frontier currently runs through the fact that the taste for what is an interesting story almost certainly remains human (because the same bottlene is happening in academic research). At least for now.","ts":1794064097,"l":44,"r":2,"c":18,"v":6803,"tp":3,"q":2,"s":38},{"id":"2119083280381383072","p":39,"t":"This can be a golden age of investigative journalism. Analyzing filings, records and public data to go deep into a company or government or whatever used to be something that took a team of journalists a long time, so was relatively rare. Now LLMs can 10x investigative research.","ts":1794063799,"l":708,"r":65,"c":157,"v":31176,"tp":2,"q":2,"s":55},{"id":"2119064880812601539","p":34,"t":"Consider that all my benchmarks are against DGX Sparks using 2200 Mhz limit on GPU frequency instead of standard ~2400. I prefer stability and lower power consumption vs pure speed.\n\nBTW now my belief that M5 Ultra can beat 4 x DGX Spark cluster are starting to shake.","ts":1794059412,"l":8,"r":0,"c":1,"v":932,"tp":1,"q":2,"s":41},{"id":"2119064491115835601","p":34,"t":"Here we are again with a new winner recipe for DeepSeek V4.1 Flash running on 4 x DGX Sparks! 🔥\n\n🥇 vLLM recipe by @christopherowen \n🥈 SGLang by @dealignai \n🥉 Tensorfold by @MiaAI_lab \n\nRecipe is here: https://github.com/christopherowen/spark-ds41f","ts":1794059319,"l":77,"r":5,"c":21,"v":6249,"tp":1,"q":2,"s":54,"ph":[["https://pbs.twimg.com/media/HWhrx7rXUAA99ao.jpg?name=orig",2400,1350],["https://pbs.twimg.com/media/HWhtQFfXYAA5V7n.jpg?name=orig",2400,1350],["https://pbs.twimg.com/media/HWhtqE_XAAAbO-m.jpg?name=orig",2094,736],["https://pbs.twimg.com/media/HWht35aWoAATteG.jpg?name=orig",1562,688]],"qt":{"id":"2118835553273553224","n":"Ivan Fioravanti","h":"ivanfioravanti","a":"https://pbs.twimg.com/profile_images/2084888225986826240/-XeooHyq_200x200.jpg","t":"We have a new winner in speed on 4 x DGX Sparks using DeepSeek V4.1 Flash 1017 lossless:\n\n🥇 SGLang image by @dealignai and @sgl_project\n🥈Tensorfold image by @MiaAI_lab and @ashxhart \n\nI will deep dive on this more tomorrow but here are some results showing SGLang winning big on prefill.\nNo cache used here, double and triple checked.\n\nModels weights are same, but SGLang version is uncensored.\n\nRecipes used:\n- SGLang https://github.com/jjang-ai/dealignai-deepseek-v4.1-flash-1017-4x-dgx-spark\n- TensorFold https://github.com/MiaAI-Lab/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold\n\nMore evals and testing on these tomorrow, together with other recipes! Let's keep pushing Local AI further!","ts":1794004736,"ph":[["https://pbs.twimg.com/media/HWedsTFWEAA0OIa.jpg?name=orig",2400,1350],["https://pbs.twimg.com/media/HWecmxTWQAAL_IX.jpg?name=orig",1130,820],["https://pbs.twimg.com/media/HWecmyHX0AAPWsA.jpg?name=orig",3316,2182],["https://pbs.twimg.com/media/HWec_jJWEAAiig2.jpg?name=orig",2528,1490]]}},{"id":"2119045613493039234","p":22,"t":"i reject PRs with love, especially if they come from folks like Nico.","ts":1794054818,"l":150,"r":2,"c":8,"v":15782,"tp":2,"q":2,"s":51,"qt":{"id":"2118964899309297897","n":"Nico Bailon","h":"nicopreme","a":"https://pbs.twimg.com/profile_images/1732784014933798912/NYfTJGk-_200x200.jpg","t":"Hot take: Pi has extensions because I wanted an undo button.\n\nLast December I opened a PR for Pi to snapshot every turn so jumping backwards in a convo could restore your file changes. Mario said no to adding it to core, then floated: \"hooks could be a way to actually extend pi.\"\n\nI opened the first hooks PR for Pi the next day. Mario rebuilt it into core based on my proposal, crediting me for the design and implementation.\n\nI began shipping many Pi hooks right away, starting with pi-rewind-hook. A few weeks later, Pi hooks grew into a full-blown extension system. I kept contributing to core and building on top of Pi: 37 extensions and 23k+ GitHub stars combined.\n\nNot bad for a rejected PR!","ts":1794035575}},{"id":"2119037007452356966","p":41,"t":"we were able to get a GB300 rack, < 1 yr commit and ready in a week for $9-10/hr\n\nfor B300s same terms for $6-7\n\nif you're shopping, sellers will insist you won't find prices that low","ts":1794052767,"l":1594,"r":22,"c":83,"v":111154,"tp":6,"q":2,"s":54},{"id":"2118973784162238748","p":41,"t":"btw OpenCode Browse supports running an agent and your browsers on different machines\n\nhad to implement this since i use a remote server, it can still orchestrate browsers on my desktop and laptop","ts":1794037693,"l":758,"r":13,"c":74,"v":75734,"tp":2,"q":2,"s":49},{"id":"2118970870979002622","p":34,"t":"Underdog team is releasing at hypersonic speed! 🚀","ts":1794036998,"l":80,"r":7,"c":8,"v":6737,"tp":8,"q":2,"s":54,"qt":{"id":"2118921155713372212","n":"Underdog AI","h":"UnderdogAI","a":"https://pbs.twimg.com/profile_images/2116283695946694656/x_eWSHUS_200x200.jpg","t":"Today we're releasing Underdog Woof, the #1 frontier TTS model\n\nIt's only 31.7 MB and beats Text-to-Speech models 1000x larger on Word Error Rate\n\nIt's small enough to run fast in your browser, phone, watch... no data centers\n\nIt’s free and open weight with Underdog CC BY-NC 4.0. Try Woof today in Underdog, your Private Personal AI: http://underdog.ai/woof","ts":1794025145,"ph":[["https://pbs.twimg.com/media/HWfpL0oaYAA76kq.jpg?name=orig",1600,900]]}},{"id":"2118898514633916828","p":41,"t":"https://x.com/opencode/status/2118895128152621153","ts":1794019747,"l":16,"r":0,"c":0,"v":7351,"tp":2,"q":2,"s":26,"qt":{"id":"2118895128152621153","n":"OpenCode","h":"opencode","a":"https://pbs.twimg.com/profile_images/1973794620233433088/nBn75BTm_200x200.png","t":"OpenCode Browse is now available\n\nit lets your agent control your real browser with your tabs and logins and extensions\n\nhttps://opencode.ai/product/browse","ts":1794018940}},{"id":"2118895958305980490","p":41,"t":"this is a must have at this point, it lets opencode (or any agent) pretty much do anything in your life\n\nall while keeping everything on device without connecting any of your accounts to some centralized cloud agent\n\nryan also made it stealthy af, cat and mouse game begins now","ts":1794019138,"l":1195,"r":10,"c":45,"v":59378,"tp":2,"q":2,"s":52,"qt":{"id":"2118895128152621153","n":"OpenCode","h":"opencode","a":"https://pbs.twimg.com/profile_images/1973794620233433088/nBn75BTm_200x200.png","t":"OpenCode Browse is now available\n\nit lets your agent control your real browser with your tabs and logins and extensions\n\nhttps://opencode.ai/product/browse","ts":1794018940}},{"id":"2118860275986399501","p":39,"t":"\"These shares imply research accelerations ranging from 1.35x in the arts, 1.36x in anthropology, and 1.41x in chemistry to 1.96x in computer science and 2.02x in statistics,\"","ts":1794010630,"l":42,"r":5,"c":5,"v":9979,"tp":3,"q":2,"s":38},{"id":"2118835553273553224","p":34,"t":"We have a new winner in speed on 4 x DGX Sparks using DeepSeek V4.1 Flash 1017 lossless:\n\n🥇 SGLang image by @dealignai and @sgl_project\n🥈Tensorfold image by @MiaAI_lab and @ashxhart \n\nI will deep dive on this more tomorrow but here are some results showing SGLang winning big on prefill.\nNo cache used here, double and triple checked.\n\nModels weights are same, but SGLang version is uncensored.\n\nRecipes used:\n- SGLang https://github.com/jjang-ai/dealignai-deepseek-v4.1-flash-1017-4x-dgx-spark\n- TensorFold https://github.com/MiaAI-Lab/DeepSeek-V4.1-Flash-1017-4x-DGX-Sparks-TensorFold\n\nMore evals and testing on these tomorrow, together with other recipes! Let's keep pushing Local AI further!","ts":1794004736,"l":88,"r":6,"c":24,"v":13414,"tp":1,"q":2,"s":55,"ph":[["https://pbs.twimg.com/media/HWedsTFWEAA0OIa.jpg?name=orig",2400,1350],["https://pbs.twimg.com/media/HWecmxTWQAAL_IX.jpg?name=orig",1130,820],["https://pbs.twimg.com/media/HWecmyHX0AAPWsA.jpg?name=orig",3316,2182],["https://pbs.twimg.com/media/HWec_jJWEAAiig2.jpg?name=orig",2528,1490]]},{"id":"2118830725738983449","p":36,"t":"People need to interact with LLMs as if they’re a representation of all humans and understand that their answers are a representation of the normal distribution of all humans (average)\n\nThat’s where anyone with an edge shines while using LLMs btw","ts":1794003585,"l":219,"r":14,"c":31,"v":8971,"tp":0,"q":2,"s":49},{"id":"2118829142607278366","p":17,"t":"our cooking show (yes, cooking show) is back!!!\n\n@aditabrm reveals the entire recipe of @reductoai pdf processing and also steak sandwich making\n\nat the end we address the serious allegations leveled by @barsisyphus \n\nmore elite food journalism coming on https://youtube.com/@latentspacepod","ts":1794003208,"l":77,"r":0,"c":30,"v":9249,"tp":2,"q":2,"s":51,"ph":[["https://pbs.twimg.com/media/HWeXlG2bcAAKkPJ.jpg?name=orig",988,1246]],"qt":{"id":"2118800336412709024","n":"Latent.Space","h":"latentspacepod","a":"https://pbs.twimg.com/profile_images/1888346877428641792/rMxtG84Z_200x200.jpg","t":"In this episode, @reductoai co-founder and CEO @aditabrm joins @allenpark to make steak sandwiches with burrata, torched peaches, and chimichurri and talk about why PDFs still break AI: how frontier models silently drop rows, why 99% accuracy can still fall short in production, what Reducto learned processing a billion pages a month, and what it would take for office work to have its Claude Code moment.\n\n00:00 Cold Open\n00:13 Meet Adit Abraham and the Viral Steak Tweet\n00:57 From Remembrall to Reducto: YC and the Pivot \n01:57 Why PDFs Are Still Hard for AI \n03:35 What AI Agents Need to Do With Documents \n04:19 Model Routing for Tricky Pages and Tables \n05:29 Build vs. Buy: When to Train Your Own Models \n07:00 Deep Extract, Frontier Models, and Missing Rows \n08:20 Why Everyone Wins Their Own Benchmark \n09:29 Reducto R1: Accuracy, Cost, and a Billion Pages a Month \n10:57 When 99% Accuracy Isn’t Enough \n13:05 Continuous Evals and Production Failures \n16:49 How Reducto Uses AI Agents Internally \n18:28 Hiring With AI and the Eight-Hour Onsite \n20:17 Early Hires and What Obsession Looks Like \n23:17 Blowtorching Peaches \n24:24 Scaling From a Small Team to Series B \n27:34 Competition and the Data Flywheel \n30:52 What Better Parsing Unlocks for Customers \n32:11 Finding a Cofounder at MIT \n34:34 The Anthropic Hackathon, Long-Term Memory, and RAG \n38:49 From Reading PDFs to Redlining Contracts \n43:19 A Claude Code Moment for Office Work\n46:13 What’s Next for Reducto: Extraction, Citations, and Cost \n47:51 Assembling the Steak Sandwich \n49:55 The Steak Allegations and Taste Test","ts":1793996340,"vd":["https://video.twimg.com/amplify_video/2118794827064397824/vid/avc1/1920x1080/EhcAvbNMSnMV0yTx.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118794827064397824/img/oosfj2VtmXEj3RGW.jpg",1920,1080]}},{"id":"2118801132646727985","p":23,"t":"These are the best demos I've ever seen, 45 seconds of pure brilliance. Finally some competition, they also seem to be in top 3 on design and vision.\n\n\"A floor-plan walkthrough generated from a map.\"\n\nCan't wait for the weights next week.","ts":1793996530,"l":243,"r":9,"c":13,"v":10871,"tp":8,"q":2,"s":54,"vd":["https://video.twimg.com/amplify_video/2118796154746056704/vid/avc1/1920x1080/S0easexx96walWwG.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118796154746056704/img/BCHSK7BUEuBe1u8a.jpg",1920,1080],"qt":{"id":"2118489527362912566","n":"Thinking Machines","h":"thinkymachines","a":"https://pbs.twimg.com/profile_images/2075275500420001792/Q9J4M1DL_200x200.jpg","t":"Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.\n\nInkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.\n\nhttps://thinkingmachines.ai/news/inkling-2-small","ts":1793922237}},{"id":"2118796604521988525","p":22,"t":"recommended viewing.","ts":1793995450,"l":121,"r":3,"c":0,"v":8405,"tp":2,"q":2,"s":49,"qt":{"id":"2118694956726440415","n":"Ben Vinegar","h":"bentlegen","a":"https://pbs.twimg.com/profile_images/1765244446650568704/zgjmJiTX_200x200.jpg","t":"🎙️ State of Agentic Coding w/ @mitsuhiko returns:\n\nIn episode #12:\n- models we're using now\n- is local desktop software dead?\n- time to hoard PS5s?\n- debating @thorstenball's \"how I work with agents\"\n\nand much more 👇","ts":1793971215,"vd":["https://video.twimg.com/amplify_video/2118527647894446080/vid/avc1/1920x1080/DMqqZ572ZN-kn4C7.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118527647894446080/img/8-PfwMO_K21Ua7VO.jpg",1920,1080]}},{"id":"2118795218099253606","p":77,"t":"We partnered with @fmf_org and its member firms on a new framework to help identify when bypassing an AI system's safeguards could enable serious cyber harm, and prioritize stronger protections.\n\n\"...This framework is an important step toward a common methodology for evaluating and comparing the severity of cyber jailbreaks, including the capabilities they unlock and how readily they can be used. Shared approaches like this can help the AI and security communities build a clearer evidence base and prioritize improvements to safeguards.\"  - OpenAI's Sasha Baker, Head of National Security Policy\n\nhttps://www.frontiermodelforum.org/updates/introducing-the-fmf-jailbreak-rating-framework-for-frontier-cyber-threats/","ts":1793995119,"l":1530,"r":124,"c":402,"v":316957,"tp":3,"q":2,"s":44},{"id":"2118792930211107115","p":75,"t":"New on our Science Blog: How AI is beginning to change scientific modeling.\n\nScientific models help researchers explain and predict real-world systems—like how a single nerve cell fires, or what happens when neutron stars collide. Recent advances have transformed how scientists fit complex models to data. But scientists still typically write the models manually, and usually need to weigh up several competing versions, each of which might explain the data in a different way.\n\nIn a guest post, Jakob Macke, Daniel Gedon, and Stefan Wahl describe how AI is starting to help with these parts of the process. By writing, running, fitting, and revising scientific models as code, AI can explore many possible model structures, and find out which explanations the data support.\n\nhttps://www.anthropic.com/research/teaching-ai-to-build-scientific-models","ts":1793994574,"l":3328,"r":365,"c":318,"v":391183,"tp":3,"q":2,"s":51},{"id":"2118789767433535935","p":23,"t":"I've had some scary experiences letting my bots go wild on my email, someone needs to build something to detect prompt injections.","ts":1793993820,"l":84,"r":2,"c":12,"v":10977,"tp":2,"q":2,"s":48,"qt":{"id":"2118789516534419558","n":"Sarah Chieng","h":"MilksandMatcha","a":"https://pbs.twimg.com/profile_images/1868428856908304384/6BROXdW-_200x200.jpg","t":"Instinct just got got. Be careful out there","ts":1793993760,"ph":[["https://pbs.twimg.com/media/HWdzn-VbMAACQcV.jpg?name=orig",1154,1458]]}},{"id":"2118782708767965238","p":77,"t":"A new way to measure how AI agents handle the routine analyses of wet-lab science—BioBasicsBench.\n\nDeveloped with Ligo, a biotech company focused on enzyme design, the benchmark draws on 100 tasks from real experiments to test whether agents can interpret raw lab data, identify errors, and recognize when the evidence doesn’t support a conclusion.\n\nAccelerating biological research starts with getting the everyday work right. BioBasicsBench shows where models need to improve, and helps us track what those gains mean for scientific progress, biosafety, and biosecurity.\n\nhttps://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science/","ts":1793992137,"l":3503,"r":262,"c":762,"v":768279,"tp":3,"q":2,"s":49,"ph":[["https://pbs.twimg.com/media/HWdtpRua4AAuNL-.jpg?name=orig",1199,890]]},{"id":"2118782509546958903","p":74,"t":"Basics of using the Union, Subtract & Intersect functions in Zoo Design Studio.\n\nhttps://youtu.be/lkbHEhmGji8","ts":1793992090,"l":5,"r":1,"c":0,"v":857,"tp":7,"q":2,"s":48},{"id":"2118763115953213831","p":41,"t":"you can run opencode server in a docker container then connect to it from the tui, desktop, etc\n\ndocker run -p 4096:4096 https://ghcr.io/anomalyco/opencode serve --hostname 0.0.0.0 --port 4096","ts":1793987466,"l":617,"r":21,"c":31,"v":33536,"tp":2,"q":2,"s":48},{"id":"2118755488108364057","p":78,"t":"While modern 2048-bit systems remain safe from these specific exploits, this work highlights a key shift.\n\nAs we enter a new era of AI-enabled math and computation, we can no longer rely on historical hardness assumptions and must continuously stress-test cryptographic standards.","ts":1793985647,"l":96,"r":4,"c":15,"v":23969,"tp":3,"q":2,"s":39},{"id":"2118740666591162831","p":45,"t":"I suspect they know this and feel the pressure. https://www.interconnects.ai/p/how-well-can-we-measure-open-model","ts":1793982113,"l":37,"r":0,"c":0,"v":5775,"tp":3,"q":2,"s":42},{"id":"2118740587666960677","p":45,"t":"All the inference startups are in a very hard place. They have the capital today to try and secure a more lasting solution — many players coming for their scalps. Need build models, their own supply, or some magic post-training solutions. \n\nWas the topic of my recent piece.","ts":1793982095,"l":401,"r":15,"c":10,"v":44273,"tp":6,"q":2,"s":56,"qt":{"id":"2118736603820761499","n":"Dylan Patel","h":"dylan522p","a":"https://pbs.twimg.com/profile_images/2113734434633768960/iNH6HMTV_200x200.jpg","t":"Why is Baseten worth more than Nebius?\nThis makes 0 sense \nVenture market valuations are nuts","ts":1793981145}},{"id":"2118732588432908369","p":34,"t":"Inkling 2 Small (594B total parameters, 17B active 👀) scoring 31 is top!","ts":1793980187,"l":60,"r":0,"c":9,"v":5932,"tp":8,"q":2,"s":52,"qt":{"id":"2118489527362912566","n":"Thinking Machines","h":"thinkymachines","a":"https://pbs.twimg.com/profile_images/2075275500420001792/Q9J4M1DL_200x200.jpg","t":"Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.\n\nInkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.\n\nhttps://thinkingmachines.ai/news/inkling-2-small","ts":1793922237}},{"id":"2118723889027051957","p":34,"t":"DecodeWall: another benchmark I'm building, focused on concurrent calls.\n\nAgain DS v4.1 Flash on 4xDGX Sparks vs 2xM3 Ultra.\n\nDGX Sparks cluster has a slight advantage in scalability. Can't wait to try something on M5 Ultra!","ts":1793978113,"l":21,"r":3,"c":10,"v":3006,"tp":1,"q":2,"s":46,"ph":[["https://pbs.twimg.com/media/HWc3yGZW0AA69Y-.jpg?name=orig",2560,1746],["https://pbs.twimg.com/media/HWc3yGUXkAAvJzy.jpg?name=orig",2510,1316],["https://pbs.twimg.com/media/HWc3yGUWkAA9MmW.jpg?name=orig",2546,558]]},{"id":"2118720876749177157","p":22,"t":"recommended viewing. @lucasmeijer is always worth your time.","ts":1793977395,"l":130,"r":9,"c":3,"v":12236,"tp":2,"q":2,"s":50,"qt":{"id":"2118679166048489608","n":"Mayank Gupta🧑🏻‍💻","h":"waghnakh_21","a":"https://pbs.twimg.com/profile_images/2085682427578732544/gkLlq-dC_200x200.jpg","t":"ambitious engineers will win in AI \n\nft. @lucasmeijer ~ indie software developer, ex-unity\n\nafter 3 months of back and forth with scheduling our pod recording, here's my experience talking to my friend @lucasmeijer.. who says, I used to be a good programmer ;)\n\nenjoy the deep insights and hot takes from Lucas on everything shaping up in AI, education and agentic coding..\n\nwhat's covered in this conversation:  \n\n00:00:00 - intro \n00:00:57 - who is Lucas and what is he building?\n00:01:38 - what did he do before joining Unity?\n00:06:03 - why did he leave the Uni of Amsterdam?\n00:07:39 - do you still need a cs degree? \n00:10:53 - programming languages for AI agents \n00:16:03 - what is the state of agentic coding today? \n00:23:37 - what is a software factory? \n00:31:09 - what is the high-value work for engineers? \n00:32:01 - are software engineering jobs going away? \n00:36:23 - building lees simpel at a hackathon\n00:38:16 - what is he building after AgentsInTheCloud? \n00:41:36 - lessons from Unity on building a company?\n00:42:40 - Lucas advice for builders learning AI\n\nfor better experience, watch the full episode on YouTube and don't forget to subscribe if you liked the content I create : https://youtu.be/D_or7cP7-Nw","ts":1793967450,"vd":["https://video.twimg.com/amplify_video/2118674036750262272/vid/avc1/1920x1080/lT_9iBdHH8-tcv8K.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118674036750262272/img/30RFf8LUruIE6vQP.jpg",1920,1080]}},{"id":"2118720645509030290","p":22,"t":"super cool!","ts":1793977340,"l":90,"r":2,"c":2,"v":9696,"tp":2,"q":2,"s":47,"qt":{"id":"2118708539610857960","n":"younes","h":"yb2698","a":"https://pbs.twimg.com/profile_images/2108460795806064640/8Q93Q0RU_200x200.jpg","t":"Harness adapters training is now live! \n\nWe are training LoRA adapters on LFM2.5 / LFM3 from @liquidai and the recent Twil-9B model from @thewebAI using RL to make them more efficient and accurate on agentic tasks! \n\nYou can follow the training going live here: https://harness-hub-harness-adapters-rl.hf.space/ \n\nhttps://huggingface.co/blog/harness-hub/harness-adapters","ts":1793974454,"vd":["https://video.twimg.com/amplify_video/2118704612056182784/vid/avc1/1920x1080/2-b5K4wzsZC6ydgT.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118704612056182784/img/Xw9MxFdVhqM8t5rU.jpg",1920,1080]}},{"id":"2118715992343789911","p":79,"t":"Stream live training runs on Hugging Face!","ts":1793976231,"l":255,"r":22,"c":92,"v":29433,"tp":0,"q":2,"s":51,"qt":{"id":"2118708539610857960","n":"younes","h":"yb2698","a":"https://pbs.twimg.com/profile_images/2108460795806064640/8Q93Q0RU_200x200.jpg","t":"Harness adapters training is now live! \n\nWe are training LoRA adapters on LFM2.5 / LFM3 from @liquidai and the recent Twil-9B model from @thewebAI using RL to make them more efficient and accurate on agentic tasks! \n\nYou can follow the training going live here: https://harness-hub-harness-adapters-rl.hf.space/ \n\nhttps://huggingface.co/blog/harness-hub/harness-adapters","ts":1793974454,"vd":["https://video.twimg.com/amplify_video/2118704612056182784/vid/avc1/1920x1080/2-b5K4wzsZC6ydgT.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118704612056182784/img/Xw9MxFdVhqM8t5rU.jpg",1920,1080]}},{"id":"2118713604539285966","p":34,"t":"AI Community wins!\n\nDGX Spark: 180.5%\nApple Silicon: 81.0%\n\n@eigenlabs @yukonresearch @darkbloomai 🔥","ts":1793975661,"l":55,"r":4,"c":4,"v":6412,"tp":1,"q":2,"s":52,"qt":{"id":"2118691730362863784","n":"Gumbii.Digital","h":"GumbiiDigital","a":"https://pbs.twimg.com/profile_images/2046926699833556992/i8H1Dr-a_200x200.jpg","t":"GM GM....Happy Friday!\n\nFinal numbers for the Qwen 3.8 27b running on the TensorFold @tensorfold_dev engine:\n\nDGX Spark: 180.5%\nApple Silicon: 81.0%\n\nHuge shout out to @ashxhart for creating such an incredible community driven engine. Also to @Alibaba_Qwen for the solid 3.8 27b framework to build off of.\n\nWe noted the anticheat change and will be looking at making less stringent moving forward.\n\nThank you again to all the participants!  These challenges would not happen without YOU!\n\n@QwenDevs @yukonresearch @pratikg @ashxhart @TheDavidTai","ts":1793970446,"ph":[["https://pbs.twimg.com/media/HWcWvXTWQAAakCw.jpg?name=orig",1139,798],["https://pbs.twimg.com/media/HWcWz0SXMAAg3lK.jpg?name=orig",1139,798]]}},{"id":"2118677527153807365","p":34,"t":"MLX-Serve RDMA for M3 Ultra PR merged! 🚀","ts":1793967060,"l":40,"r":0,"c":2,"v":3792,"tp":1,"q":2,"s":50,"qt":{"id":"2118568973704855804","n":"David Dalcu","h":"ddalcu","a":"https://pbs.twimg.com/profile_images/2075639671766274048/4VbirihF_200x200.jpg","t":"MLX-Serve 26.11.2 is out\n⭐️ https://github.com/ddalcu/mlx-serve/releases\nMany ways to run Deepseek from ~80gb - 512 ram.\n@ivanfioravanti 's RDMA code is in for the lossless quant + Massive M3 Speedups\n@Beamsters1 SushiSeek STQ\n@TheDavidTai 's EXL3 quants\n@UnslothAI or GSQ-RCO GGUF Quants \n\nThank you:\nelkaix, @fedeandri, @garnetlyx, @hyspace, jondestoppeleire, lborloz, pcoppinger, sjordan0228, slava-kudzinau, @STRML_ , toxeh, trigger2k20, @yoyo930021","ts":1793941179,"vd":["https://video.twimg.com/amplify_video/2118565643515551744/vid/avc1/1920x1080/a_WxDr_77obrXEz7.mp4?tag=29","https://pbs.twimg.com/amplify_video_thumb/2118565643515551744/img/RhzJp7OVaEPU6-YY.jpg",1920,1080]}},{"id":"2118674929797517492","p":34,"t":"Months of digging into MoE training condensed int The Ultimate MoE Training Handbook! \nWeekend reading slot booked!","ts":1793966440,"l":31,"r":1,"c":1,"v":3219,"tp":0,"q":3,"s":57,"qt":{"id":"2118412817116369254","n":"Nouamane Tazi","h":"Nouamanetazi","a":"https://pbs.twimg.com/profile_images/1523789465973633024/TnwBVpmT_200x200.jpg","t":"📖 The Ultimate MoE Training Handbook 📖\n\nWe spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.\nhttps://huggingface.co/spaces/huggingface/moe-blogpost","ts":1793903948,"vd":["https://video.twimg.com/amplify_video/2118412801958199296/vid/avc1/1280x720/LXqlteM6gb4sQ4KA.mp4?tag=14","https://pbs.twimg.com/amplify_video_thumb/2118412801958199296/img/_bn2vtFt2Xjha11G.jpg",1920,1080]}},{"id":"2118659381856550950","p":34,"t":"Let's keep pushing both Apple Silicon and DGX Spark!","ts":1793962734,"l":47,"r":1,"c":3,"v":3474,"tp":1,"q":2,"s":51,"qt":{"id":"2118488485963485662","n":"Ash Hart","h":"ashxhart","a":"https://pbs.twimg.com/profile_images/2076633444411494400/InvR8zXt_200x200.jpg","t":"I've taken matters into my own hands with @Apple's Metal. 🔥\n\nI'm building my own custom Metal kernel for TensorFold, starting with 4-bit inference on M5.\n\nIt'll set me back a little while I get this built and tested properly, but what I've achieved locally with (mlx_tf) already feels like a real breakthrough.\n\nThe M3 Ultra has 819 GB/s of memory bandwidth. The M5 Ultra has 1.2 TB/s. I look at that hardware and think there's still a lot more decode performance to unlock.\n\nWhy should the Sparks have all the fun... Watch this space 👀","ts":1793921989,"ph":[["https://pbs.twimg.com/media/HWZiDENXYAAHR_P.jpg?name=orig",1200,674]]}},{"id":"2118654722194207159","p":22,"t":"neat","ts":1793961623,"l":133,"r":4,"c":8,"v":19056,"tp":2,"q":2,"s":50,"qt":{"id":"2118373808692437403","n":"Joël Niklaus","h":"joelniklaus","a":"https://pbs.twimg.com/profile_images/1468993727926620161/I1lBhFJM_200x200.jpg","t":"I burned a shitload of compute so you don't have to. Releasing the Harness Leaderboard.\n\nWhich harness should you run your open model in? I ran 28,849 trials to find out.\n\nThe leaderboard compares nine coding-agent harnesses on open models. In each comparison only the harness changes: the model, provider and reasoning effort stay fixed, and every task gets three attempts.\n\nThe trials ran GLM-5.3-Flash, DeepSeek-V4.1-Flash and Flint on Terminal-Bench 2.1 and DeepSWE, for $4,746 of inference.\n\nAveraged over the three models, mini-swe-agent (@KLieret), a bash-only agent, solves 51.4% of tasks and @openclaw 27.3%. Four others are statistically tied with mini-swe-agent.\n\nOne of them is Pi (@pidotdev), at 49.4% for $0.29 per solved task, about 45% less than mini-swe-agent's $0.53.\n\nThe best harness depends on the model: mini-swe-agent on GLM-5.3-Flash, Hermes Agent (@NousResearch) on DeepSeek-V4.1-Flash and Kimi Code (@Kimi_Moonshot) on Flint.\n\n@opencode costs $3.35 per solved task, 11.7 times as much as Pi.\n\nSee Harness Hub's Harness Lens for what the nine harnesses do differently. It documents the prompts, tools and compaction of 34 coding-agent harnesses and links every claim to its source.\n\nLeaderboard: https://huggingface.co/spaces/joelniklaus/harness-leaderboard\nHarness Lens: https://huggingface.co/spaces/harness-hub/harness-lens","ts":1793894648,"vd":["https://video.twimg.com/amplify_video/2118373780292874240/vid/avc1/1920x1080/wQA81GN6E0eWLY51.mp4?tag=16","https://pbs.twimg.com/amplify_video_thumb/2118373780292874240/img/XlN2NzAPDJpW299Q.jpg",1920,1080]}},{"id":"2118634107672592531","p":34,"t":"DeepSeek v4.1 Flash 1017 Original Weights test in progress! Working on a comparison:\n- 4 x DGX\n- 2 x M3 Ultra\n- 1 x M3 Ultra\n\nSpoiler: I don't have an M5 Ultra yet, but I bet it wins against the 4 x DGX. We'll discover this soon!","ts":1793956708,"l":53,"r":4,"c":9,"v":6704,"tp":1,"q":2,"s":52},{"id":"2118620978632904721","p":21,"t":"I get questions like this a lot so I'll go on a quick rant here:\n\nIf you only have one AI sub, one computer, and no desire to code on your phone, T3 Code is not that big an improvement.\n\nAs soon as ANY of those things change, T3 Code is the legit only reasonable option.\n\nI hope that changes, but for now you're taking a massive productivity hit using literally anything else 🙃","ts":1793953578,"l":1142,"r":24,"c":178,"v":103666,"tp":2,"q":2,"s":50,"ph":[["https://pbs.twimg.com/media/HWbaQzmbIAAQQJo.jpg?name=orig",966,1320]]},{"id":"2118564175194849388","p":36,"t":"Setting up a Mac mini with Apple Messages that uses a local model to help the wife troubleshoot and add movies & shows to Plex was somehow among the top 3 thing I did in the past couple years lmao","ts":1793940035,"l":124,"r":2,"c":22,"v":9643,"tp":1,"q":2,"s":46},{"id":"2118541829176226015","p":41,"t":"gpui is really great considering it was pre-ai and it tackled one of the most annoying problems ever\n\nbut post-ai it wasn't able to meet the bar we wanted\n\nwe probably need to do some work to explain rune better but also just go look at how nice hex is on all platforms","ts":1793934707,"l":1387,"r":15,"c":56,"v":96428,"tp":5,"q":2,"s":53},{"id":"2118521863173451804","p":21,"t":"Day 34/","ts":1793929947,"l":411,"r":7,"c":25,"v":26186,"tp":2,"q":2,"s":44,"qt":{"id":"2118521476542558265","n":"Julius","h":"jullerino","a":"https://pbs.twimg.com/profile_images/1526143235067985922/DmFx1k13_200x200.jpg","t":"You can now run local transcription models on your @t3dotcodes environments and use it as a dictation hub for any client! Choose among an abundance of models, filter through which best suits your needs, dictionary for words the models don't get right for your voice, and optional customizable post-processing!\n\nTry it out and send feedback!","ts":1793929854,"ph":[["https://pbs.twimg.com/media/HWZ-zziakAAR9or.jpg?name=orig",3600,2260],["https://pbs.twimg.com/media/HWZ_uXaaQAAtFfl.jpg?name=orig",3600,2260]]}}],"shorts":[["vN7D2hmUJvI","Kepler’s Impossible Equation",1300000,2,4,3,82],["BuljHijAK4o","Imaginary Numbers and Astronomy",861000,2,4,3,79],["35IpOK-WaNA","The efficient compute frontier.",5400000,2,3,2,78],["KpmrQdaQjGg","The Ptolemaic Model",6300000,2,4,2,78],["m_sBIczld_8","Building an Elevator to Outer Space – Adam Brown",22000000,14,7,2,78],["v05HU2nBrOs","When support structures fail, try this!  #3dprinting tip",5700000,70,7,2,77],["Sc4595WJHCc","The Most useful Curve In mathematics",1300000,2,4,2,74],["KIx6FrcrhLE","$1,000,000 Ancient Scroll Challenge",7300000,14,0,2,74],["pLHWn5SQvLM","The Era of Easy AI Progress Is Ending - Ilya Sutskever",1700000,14,3,3,73],["IfmYtKLjqq0","Alexnet in 60s.",841000,2,0,2,71],["tfBT22RWYww","What is the relationship between LLMs, Reasoning Models, and Agent Harnesses?",6100,5,2,3,71],["30X4KfOcqj8","Imaginary Numbers are Real.",193000,2,4,3,70],["uVS88gnaxcg","Meet Claude Fable 5.1",149000,75,8,3,70],["-auPuUxNWSc","How Kepler handled his impossible equation.",156000,2,4,3,69],["0aG4cSfFvC0","The Attention Mechanism for Large Language Models #AI #llm #attention",15000,9,0,3,69],["Y_adWrLAW7A","Noam Chomsky - Science Vs Engineering",74000,13,3,2,69],["1ajWYL7Nddo","Introducing RAG 2.0: Agentic RAG + Knowledge Graphs (FREE Template)",170000,20,2,2,69],["kR830P78mgI","This is an Anti Oloid",1500000,70,4,2,69],["FIgIVO9GvlA","Why are atomic weights not round?",520000,2,4,2,68],["wtLQUjfomig","The Exponential Distribution",163000,3,4,3,68],["VjGeZRvcMOo","Agentic AI vs LLMs: Why an Agent Doesn't Stop at the Answer",115000,7,2,2,68],["Tp_mnXrB9Pc","FreeCAD Quick Tip : Using Construction Geometry To Centered Sketch Geometry in Sketcher",101000,73,7,2,68],["FJtFZwbvkI4","How word vectors encode meaning",3400000,1,0,3,67],["AZ0pNUF1kAc","How AI Coding Agents Are Changing Software Development",100000,7,2,2,67],["3lQXYNUyv5o","Why AI Agents Matter Now with Stanford's Azalia Mirhoseini and Aakanksha Chowdhery",98000,7,2,2,67],["aIvHf8vsWBM","Why Vibe Coding Fails - Ilya Sutskever",2300000,14,2,2,67],["qsjIdzy8DT4","3 New PCs, One Giant AI Model… This Shouldn’t Work",641000,24,1,2,67],["aIqi8kx2qAw","150YO Expanding Pulley Mechanism",1000000,70,7,2,67],["IWzWTTy3ng4","Does a wobbly table RUIN print quality?",999000,70,7,2,67],["xIMlJUwB1m8","The medical test paradox (well \"paradox\")",11000000,1,4,2,66],["qoogokA-dsU","Quick! Can you solve for E?",354000,2,4,2,66],["UJpzv4jDpoU","Creating the oval.",403000,2,4,2,66],["voKKTN1ocvs","Neel Nanda on mechanistic interpretability #artificialintelligence #gpt",13000,13,3,3,66],["AeIsIVgjP4A","Real Intelligence is Continual Learning - Ilya Sutskever",573000,14,3,3,66],["-CoxTFc6Lg4","Fixing the Sun.",308000,2,4,2,65],["953NHzFtGHc","Likelihood vs Probability",108000,3,4,3,65],["IcFX0JJcumo","Different Tokens, Different Compute in LLMs",2600,5,0,3,65],["GgMfuD4JsJo","Enigma Machines: The WW2 Codebreaking Race - Sarah Paine",1900000,14,5,2,65],["fmEGBwllKDU","What data is needed to train an AI model like Deepseek R1? #ai #coding #podcast",10000,17,0,2,65],["tZjmK4qdpe0","NVIDIA RTX 5080 Ollama test",502000,24,1,2,65]]}