SWE-chat by @joabaum is a living dataset of human interactions with AI coding agents. 230K prompts from real developers collected in the wild are now available on HuggingFace 🚀 huggingface.co/datasets/SALT…

Some interesting and perhaps unexpected properties of tokenization for language models!

Most SWE benchmarks tell an agent exactly what to fix + where to look. Imagine if an agent could instead *proactively* surface and fix issues themselves (before a developer even runs into them!) We evaluate this capability at scale. Introducing SWE-sweep! Led by 🐐 @KLieret

This work directly tackles the largely unexplored problem of submission overload at conferences. It argues that technical governance solutions can help everyone build trust and faith in our system, and in turn allows building on each other's contributions. Check it out!

Bringing up the software stack for a new chip has always been a major bottleneck. That's about to change as we enter the era of AI as a compiler! The model directly translates source code (e.g., Triton) to low-level assembly (e.g., PTX), and a verifier checks functional correctness through symbolic / numeric analysis, as well as other properties such as race conditions and deadlocks! No need for manually constructed layers of IR and DSLs. Speedups over Triton baselines on B200: FlashAttention (1.37x) Mamba2 state forward (1.10x) Great work led by @cos_francois, in collaboration with Charly Castes and Thomas Bourgeat from EPFL! More to come soon!

Kinda crazy to get an Oral for your first paper - congratulations @alexxx_zzz6825 !! 🥳🎉🎊

Sorry for delay! Here is my work on zero order (ZO) optimization. We achieve SOTA for ZO methods on pretraining controlling for compute and parameters. ZO methods struggle to improve loss as model size grows because relative gradient variance increases linearly with the number of perturbed parameters, inhibiting large model training. All previous methods innovate on the optimizer, but adopt architectures designed for backprop. Instead, we design an architecture with ZO in mind, that caps the gradient variance as you increase model size. Introducing SOMA (Sharded Optimization Mixture of Assemblies). SOMA shards the model into tiny experts, and trains each expert independently, on disaggregated GPUs, without communicating gradients, activations or optimizer state during training. Inspired by the thalamus and cortical columns, each expert specializes on a subset of the train set based on a fixed router. At isocompute and isoparams, SOMA achieves lower loss vs. all monolithic ZO methods tested (e.g. EGGROLL, more perturbations w Vanilla SPSA, etc). ZO w SOMA continues to improve at larger model sizes. Additional inference benefits of this architecture, we can select top-k active experts over N trained experts to reduce inference compute by O(N/k), allowing us to tradeoff inference flops for accuracy wo retraining.

We found that LMs, large or small, text-only or multimodal, are susceptible to a shared problem -- 𝗼𝘃𝗲𝗿𝗰𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗼𝗯𝗹𝗶𝘃𝗶𝗼𝗻 𝘁𝗼 𝗱𝗲𝘁𝗮𝗶𝗹𝘀. When presented with a scenario that is uncommon with overwhelming pretraining data that encourages a shortcut, they take the shortcut confidently all the time. A picture that looks like a dog? Four legs. Location for the last Olympics? Paris. A question that looks like it needs detailed answers? Start a bullet point list. Simple finetuning won't cure this problem, especially when the baked-in behavior is so overwhelmingly represented in training -- but a simple modified preference learning technique would solve it perfectly. I'll be presenting this technique, Abductive Preference Learning, next Thursday at #COLM2026, come say hi! Joint work with Yijin Ni and @simon_ycl

Mercury Voice is here, bringing diffusion speed to agentic voice. Excited to see what developers build with it.

Looped transformer meets agent swarm. Our new @Neurips2026 paper shows how to loop compute across multiple agents. Different LLMs recursively share latent thoughts w/ each other. Better performance with up to 75% fewer tokens. Great work led by @Jiaru_Zou w/ awesome collaborators.

Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI. Perhaps counterintuitively, I think it's a great time to do so. I wrote up some thoughts on this here: web.mit.edu/phillipi/www/wri…

So excited to welcome @theworldlabs and @drfeifei to the @AMD family! I’ve always been a huge fan of Fei-Fei and her pioneering research in AI. Together, we’ll combine World Labs’ deep expertise in AI and world models with AMD’s compute leadership to power the future of AI and strengthen the open AI ecosystem. Can’t wait for all we’ll accomplish!

Excited to share our #NeurIPS2026 paper on benchmarking hallucinations in LLMs! We take a fresh world model-centric approach to measuring hallucinations in several environments including terminal, grid worlds, and chess. Thanks to the @DegenAI_Labs team for their hard work! Also grateful to collaborators at @Stanford, @StanfordAILab, @stanfordnlp, @CarnegieMellon, @LTIatCMU, @OhioState, and @OhioStateCSE.

Congratulations to AIMI-affiliated faculty Ehsan Adeli, PhD, and colleagues on a new $25M @NIHAging grant to establish a Stanford center exploring how AI can help personalize care for people with Alzheimer’s disease and related dementias. 🔗 Learn more: stan.md/4hhURQA

Life update - I left OpenAI two months ago. I'm very thankful to have had the chance to contribute to agi and work with great teammates (miss ya'll ❤️). I'm now excited to start working on some of the problems I think will be important after agi...specifically robotics! 🤖🤖

PSA: While our names are similar, @mmitchell_ai and I are actually two completely different people 😂

1/10 Can we build a world model of human health? I'm thrilled to announce my first senior author paper: HealthFlux! Simulating how health evolves over a lifetime to find ways to prevent disease. doi.org/10.64898/2026.09.19.…

What would pretraining with zero real data look like? Can a randomly initialized model learn to generate all of its training data entirely through self-play? Find out below! :)

Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks. With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks. We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed. Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size. 📄 Blog: contrastive-lm.notion.site 💻 Code: github.com/Contrastive-LM/CL… 🗣️ Discord: discord.gg/5dAQEDJBs 🤗 Data & Models: huggingface.co/Contrastive-L… More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵