AI300
  1. Omar Khattab / MIT CSAIL Asst professor
    1.
    Late Interaction post-trains RLM-Qwen3-8B at tiny scale into the first natively recursive LM, releases it, and shows promise on unseen tasks.
    LLM score 85 · about 12 hours ago
  2. Alex Zhang / MIT CSAIL PhD
    2.
    Released RLM-Qwen3-8B, the first natively recursive language model (at tiny scale), post-trained on just ~1000 RLM trajectories from unrelated domains.
    LLM score 85 · about 13 hours ago
  3. alphaXiv
    3.
    Paper introduces Self-Distillation Fine-Tuning (SDFT) for on-policy continual learning from expert demonstrations without explicit reward inference or engineering.
    LLM score 85 · about 15 hours ago
  4. Andrew Lampinen / Research Scientist at DeepMind
    4.
    New paper shows language model representations of factuality evolve dramatically in edge-case conversations on consciousness or delusions.
    LLM score 85 · about 16 hours ago
  5. Cameron Wolfe / Researcher at Netflix
    5.
    Trinity Large is a sparse 400B-A13B MoE with 256 experts (4 active per token), chosen for a ~30-day training sprint that also yields blazing speed plu
    LLM score 85 · about 21 hours ago
  6. Ben Burtenshaw / Hugging Face Researcher
    6.
    Claude teaches open models to write CUDA kernels via agentic capability transfer, detailed in this blog post.
    LLM score 85 · 1 day ago
  7. alphaXiv
    7.
    New interpretability idea 'Patterning': reverse-engineer training data that induces desired model circuits/algorithms by measuring internal sensitivit
    LLM score 75 · 1 day ago
  8. Boris Cherny / Creator of Claude Code
    8.
    @charles_irl In case it’s not clear in the docs: - Ancestor https://t.co/v4FOLUBHz9’s are loaded into context automatically on startup - Descendent https://t.co/v4FOLUBHz9’s are...
    LLM score 20 · 2 days ago
  9. alphaXiv
    9.
    Paper shows LLM in sandbox with browsing, file ops, and code execution unlocks general agentic intelligence and boosts non-code performance—sans train
    LLM score 85 · 2 days ago
  10. Ethan Shen / Ai2 Researcher
    10.
    GA uses 4.5 Air as teacher; SERA-32B uses GLM 4.6. GLM 4.7 Flash flops at 32K/64K context, possibly too reliant on long contexts.
    LLM score 85 · 2 days ago
  11. Tim Dettmers / Research Scientist at Ai2
    11.
    his work was mostly the genius of Ethan Shen. He started without a background in coding agents and is now one of the most knowledgeable people in the coding agent area. He has all...
    LLM score 70 · 3 days ago
  12. Tim Dettmers / Research Scientist at Ai2
    12.
    Our method became so efficient (26x vs RL; 57x vs other synth gen), that we could easily generate 1000s of trajectories for a single repo. This makes the coding agents very...
    LLM score 85 · 3 days ago
  13. Tim Dettmers / Research Scientist at Ai2
    13.
    This is very impactful: you can now distill frontier performance into small models that are specialized to private repositories. Companies can quickly and cheaply train on their...
    LLM score 75 · 3 days ago
  14. Tim Dettmers / Research Scientist at Ai2
    14.
    From there we could do a massiv amounts of experiments and really understand what matters for training coding agents. The most important insights came from carefully evaluating...
    LLM score 80 · 3 days ago
  15. Cameron Wolfe / Researcher at Netflix
    15.
    Standard RL excels at continual learning, suggesting it's closer for LLMs than thought.
    LLM score 65 · 3 days ago
  16. Ethan Shen / Ai2 Researcher
    16.
    SWE-Bench demands multiple seeds for statistically significant results; single-seed scores, often reported, are meaningless.
    LLM score 95 · 3 days ago
  17. Ethan Shen / Ai2 Researcher
    17.
    We experiment with mixing SERA's first and second rollout data, finding they complement each other to boost performance in data-constrained regimes de
    LLM score 85 · 3 days ago
  18. Ethan Shen / Ai2 Researcher
    18.
    Best data comes from truncated trajectories that just barely exceed the context limit.
    LLM score 85 · 3 days ago
  19. Niklas Muennighoff / AI Researcher at Stanford
    19.
    Community-built open benchmarks work really well, e.g., Terminal-Bench, HLE, MMTEB. Anyone can improve them by adding tasks without breaking others. If models had the same...
    LLM score 80 · 3 days ago
  20. Noam Brown / OpenAI Research Scientist
    20.
    Had to cut this one for space: 2019: AI can't create art—creativity is uniquely human
    LLM score 20 · 3 days ago
  21. Andrej Karpathy / AI researcher
    21.
    @0xabi96 It feels like I’m cheating. Which is a very weird feeling to have. It takes a while to unpack. It’s because some code that used to be a point of pride and high IQ and...
    LLM score 70 · 3 days ago
  22. Andrej Karpathy / AI researcher
    22.
    @ChiragLathiya The nearest neighbor really is some kind of a junior engineer. Its ideas about what experiments to run on e.g. nanochat have been surprisingly bad, but its...
    LLM score 80 · 3 days ago
  23. Andrej Karpathy / AI researcher
    23.
    @jeremytwei Love the word "comprehension debt", haven't encountered it so far, it's very accurate. It's so very tempting to just move on when the LLM one-shotted something that...
    LLM score 30 · 3 days ago
  24. Andrej Karpathy / AI researcher
    24.
    @airesearch12 💯 @ Spec-driven development It's the limit of imperative -> declarative transition, basically being declarative entirely. Relatedly my mind was recently blown...
    LLM score 60 · 3 days ago
  25. Andrej Karpathy / AI researcher
    25.
    A few random notes from claude coding quite a bit last few weeks. Coding workflow. Given the latest lift in LLM coding capability, like many others I rapidly went from about 80%...
    LLM score 20 · 3 days ago
  26. Noam Brown / OpenAI Research Scientist
    26.
    1987: AI can't win at chess—planning is uniquely human 1997: AI can't win at Go—intuition is uniquely human 2016: AI can't win at poker—bluffing is uniquely human 2023: AI...
    LLM score 70 · 3 days ago
  27. Yoshua Bengio
    27.
    Had a great discussion on AI's major societal impacts and emerging risks on the @RestIsPolitics podcast this past November. Thank you @RoryStewartUK and @matthewclifford for...
    LLM score 20 · 3 days ago
  28. Jonathan Ross / TPU Creator
    28.
    Success in the Information Age was about being able to answer questions. Success in the AI Age will be about being able to ask the right questions.
    LLM score 20 · 3 days ago
  29. Ben Burtenshaw / Hugging Face Researcher
    29.
    this is a blog post on claude + llama.cpp https://t.co/yej6WsNnQA
    LLM score 20 · 4 days ago
  30. Yoshua Bengio
    30.
    Merci à @HugoDecrypte pour l'invitation à faire un tour d'horizon des risques de l'IA ainsi que des solutions nécessaires pour se diriger vers une meilleure trajectoire. Depuis...
    LLM score 20 · 4 days ago
More