In my view, SWE-bench and T-bench are the few benchmarks that have a good signal in terms of how much progress we make with models. This model performs as well as Qwen3 Coder and is only 10% worse than GPT-5, while also being a general-purpose LLM rather than code-specialized.

Shower of thoughts: Instead of keeping your Twitter/𝕏 payout, direct it towards a "PayoutChallenge" of your choosing - anything you want more of in the world! Here is mine for this round, combining my last 3 payouts of $5478.51: It is imperative that humanity not fall while AI ascends. Humanity has to continue to rise, become better alongside. Create something that is specifically designed to uplift team human. Definition intentionally left a bit vague to keep some entropy around people's interpretation, but imo examples include: - Any piece of software that aids explanation, visualization, memorization, inspiration, understanding, coordination, etc... - It doesn't have to be too lofty, e.g. it can be a specific educational article/video explaining something some other people could benefit from or that you have unique knowledge of. - Prompts/agents for explanation, e.g. along the lines of recently released ChatGPT study mode. - Related works of art This challenge will run for 2 weeks until Aug 17th EOD PST. Submit your contribution as a reply. It has to be something that was uniquely created for this challenge and would not exist otherwise. Criteria includes execution, leverage, novelty, inspiration, aesthetics, amusement. People can upvote submissions by liking, this "people's choice" will also be a factor. I will decide the winner on Aug 17th and send $5478.51 :)

Some updates 🚨 I finished my Ph.D at @uwcse in June 2025! After a year at AI2 as a Research Scientist, I am joining CMU @LTIatCMU & @mldcmu (courtesy) as an Assistant Professor in Fall 2026. The journey, acknowledgments & recruiting in 🧵

Announcing our early work on FP4 inference for LLMs! - QuTLASS: low-precision kernel support for Blackwell GPUs - FP-Quant: a flexible quantization harness for Llama/Qwen We reach 4x speedup vs BF16, with good accuracy through MXFP4 microscaling + fused Hadamard rotations.

Tokenization is just a special case of "chunking" - building low-level data into high-level abstractions - which is in turn fundamental to intelligence. Our new architecture, which enables hierarchical *dynamic chunking*, is not only tokenizer-free, but simply scales better.

The biggest dataset of human written GPU Code all open-source? 👀 YES Please! We at @GPU_MODE have released around 40k 🚀 human written code samples spanning Triton, Hip and PyTorch and it's all open on the @huggingface Hub. Train the new GPT to make GPTs faster ⚡️ Link below ⬇️

I really like this result: an elegant framing and solution to significantly improve length generalization in recurrent models at large (RNNs/SSMs/linear attention/etc). This has significant implications for the problems architecture researchers should focus on, IMO

I should probably announce that a few months ago, I joined @scale_AI to lead the Safety, Evaluations, and Alignment Lab… and today, I joined @Meta to continue working on AI alignment with @summeryue0 and @alexandr_wang. Very excited for what we can accomplish together!

What will software development look like in 2026? With coding agents rapidly improving, dev roles may look quite different. My current workflow has changed a lot: - Work in github, not IDEs - Agents in parallel - Write English, not code - More code review Thoughts + a video👇

📢Now open, Gemma 3n weights & it is natively flexible, first of its kind, thanks to MatFormer🪆 Any model between E4B & E2B with ZERO training near Pareto -- we found a bunch! Find a better E3B than what we released, I will send you a 🪆😉 Find the colab for extraction 🧵👇🪆