AI300
  1. Leandro von Werra / Hugging Face Head of Research
    1201.
    Asked the web instead of the code agent to make a plot of some data and it came up with a creative use of the website screenshot tool: https://t.co/AhOnpTwBQp
    LLM score 82 · 4 months ago
  2. Chelsea Finn / Physical Intelligence Cofounder
    1202.
    Video gen models make pretty videos, but lack physical accuracy›
    LLM score 92 · 4 months ago
  3. Lucas Beyer / Meta Researcher
    1203.
    So, when I have GitHub Copilot point out a typo in my PR, and I'm lazy, I click "Commit suggestion" and it makes a commit.›
    LLM score 82 · 4 months ago
  4. Boris Cherny / Creator of Claude Code
    1204.
    Sonnet 4.6 is now live in Claude Code.›
    LLM score 85 · 4 months ago
  5. Boris Cherny / Creator of Claude Code
    1205.
    Loving Sonnet 4.6 in Cowork -- great balance of capability, speed, and token efficiency https://t.co/ZQErYmBgtE
    LLM score 65 · 4 months ago
  6. Lucas Beyer / Meta Researcher
    1206.
    I usually look at which benches the small model surpasses its previous big brother, if it's only few, i think it gives a hint as to what they focus on.›
    LLM score 82 · 4 months ago
  7. Eric Jang / ex VP of AI at 1X Robotics
    1207.
    Incredibly impressive. We are in the early days of seeing how much there is yet to be unlocked by the humanoid form factor https://t.co/4dXbAPD48x
    LLM score 25 · 4 months ago
  8. Lewis Tunstall / Hugging Face Researcher
    1208.
    Haha, I tried the car wash riddle on QED-Nano and not only did it get it right, but it made the creative suggestion to avoid driving altogether!›
    LLM score 82 · 4 months ago
  9. John Carmack
    1209.
    The glory work of GPU scheduling is in the frontier data centers with hundreds of thousands of GPUs, but a lot of research work is done with single GPU jobs on modest clusters, and the scheduling leaves much to be desired.›
    LLM score 85 · 4 months ago
  10. alphaXiv
    1210.
    Making LLMs truly learn from its experience "Experiential Reinforcement Learning (ERL)"›
    LLM score 92 · 4 months ago
  11. Cameron Wolfe / Researcher at Netflix
    1211.
    Rubric-based RL is one of the most active topics in AI research because it extends the benefits of large-scale RL training to non-verifiable domains.›
    LLM score 95 · 4 months ago
  12. Ben Burtenshaw / Hugging Face Researcher
    1212.
    bold small model release from @Cohere_Labs to add region focused versions of their Tiny Aya model.›
    LLM score 85 · 4 months ago
  13. Wenting Zhao / Alibaba Qwen Researcher
    1213.
    For Qwen-3.5, We took the training recipe from Qwen3-Coder-Next and scale up model parameters, and this results in a much stronger coding agent.›
    LLM score 92 · 4 months ago
  14. Andrew White / Edison Scientific Cofounder
    1214.
    I've been working on and off over the last year to update the book.›
    LLM score 85 · 4 months ago
  15. Sergey Levine / Physical Intelligence Cofounder
    1215.
    VLAs can enable vehicles to better handle complex edge cases: a VLM can "think through" a complex interaction, deduce a common sense behavior, and then a VLA can carry that out to maintain safe(r) behavior even in unusual situations.›
    LLM score 92 · 4 months ago
  16. Sander Dieleman / DeepMind Research Scientist
    1216.
    This post provides a comprehensive overview of the intersection of generative modelling and representation learning (think REPA, VA-VAE, RAE).›
    LLM score 92 · 4 months ago
  17. Lucas Beyer / Meta Researcher
    1217.
    Here's another prompt I quite like using, I call it `lbreview`, because while there are tons of "review pls" prompts, this one is mine, and I like it!›
    LLM score 85 · 4 months ago
  18. Omar Khattab / MIT CSAIL Asst professor
    1218.
    for all i can tell, what makes reasoning models work so well is that, at sufficient pretrain/RL scale, every relevant next-reasoning step can actively be “softly contaminated” and composed up›
    LLM score 92 · 4 months ago
  19. alphaXiv
    1219.
    Discrete Diffusion just got a huge upgrade! with "Categorical Flow Maps", it is now much faster + capable of test time inference.›
    LLM score 92 · 4 months ago
  20. Lucas Beyer / Meta Researcher
    1220.
    Back in the day when we figured out that large-scale pretraining seems to be a great thing, I then very quickly decided to spend a significant amount of time and effort towards removing all our downstream transfer/eval task's images from our pretraining data.›
    LLM score 92 · 4 months ago
  21. Andrej Karpathy / AI researcher
    1221.
    I think it must be a very interesting time to be in programming languages and formal methods because LLMs change the whole constraints landscape of software completely.›
    LLM score 92 · 4 months ago
  22. Sander Dieleman / DeepMind Research Scientist
    1222.
    Training continuous diffusion models for discrete data with cross-entropy is neat (😎), but training flow maps with cross-entropy is even neater! Nice find! https://t.co/ZV1TDTykzY
    LLM score 92 · 4 months ago
  23. Omar Khattab / MIT CSAIL Asst professor
    1223.
    Though bash is a completely valid REPL, the amount of time coding agents lose during experimentation because they iterate on scripts instead of a Jupyter-like in-memory REPL is basically dumb.›
    LLM score 92 · 4 months ago
  24. Omar Khattab / MIT CSAIL Asst professor
    1224.
    Not saying this is a hard one, but: - ChatGPT 5.2 Thinking indeed struggled for me›
    LLM score 82 · 4 months ago
  25. Omar Khattab / MIT CSAIL Asst professor
    1225.
    ok looks like we’re past peak monthly downloads for this ancient 2021 ColBERTv2 model checkpoint›
    LLM score 92 · 4 months ago
  26. Omar Khattab / MIT CSAIL Asst professor
    1226.
    Not saying this is a hard one, but: - ChatGPT 5.2 Thinking indeed struggled for me›
    LLM score 82 · 4 months ago
  27. Sergey Levine / Physical Intelligence Cofounder
    1227.
    If we train VLAs to respond to diverse multimodal prompts, then we can steer them better: [grasp the carrot]/[move to x,y,z]/[put the carrot on the plate].›
    LLM score 92 · 4 months ago
  28. Ben Burtenshaw / Hugging Face Researcher
    1228.
    so 'harness' has increased a lot in the ai vocab over the last few months.›
    LLM score 85 · 4 months ago
  29. Cameron Wolfe / Researcher at Netflix
    1229.
    Writeup on Rubric-Based RL is out now: https://t.co/io6zEeeEAZ›
    LLM score 92 · 4 months ago
  30. Thomas Wolf / Hugging Face Cofounder
    1230.
    Shifting structures in a software world dominated by AI.›
    LLM score 85 · 4 months ago
More