Jeffrey Li

A statement on the comments from Secretary of War Pete Hegseth. anthropic.com/news/statement…

OpenThoughts is going to be an Oral Presentation at ICLR! It's my first oral presentation so super excited! See y'all in Brazil! :)

We're releasing🍨Gelato-30B-A3B, a state-of-the-art computer grounding model that delivers immediate performance gains for computer-use agents! Trained on our open-source🖱️Click-100k dataset, Gelato achieves 63.8% on ScreenSpot-Pro and 69.1% on OS-World-G. It outperforms specialized models like GTA1-32B and VLMs ~8× its size like Qwen3-VL-235B. (1/N) 🧵

Today, we’re announcing the next chapter of Terminal-Bench with two releases: 1. Harbor, a new package for running sandboxed agent rollouts at scale 2. Terminal-Bench 2.0, a harder version of Terminal-Bench with increased verification

> we've hit a data wall > pretraining is dead Is it? Today we are releasing 📄 FinePDFs: 3T tokens of new text data for pre-training that until now had been locked away inside PDFs. It is the largest permissively licensed corpus sourced exclusively from PDFs.

🚨Deadline extended to September 1 for #NeurIPS 2025 CCFM Workshop. You now have 10 days more to submit your work!

📢Submissions are now open for #NeurIPS2025 CCFM workshop. Submission deadline: August 22, 2025, AoE. Website: sites.google.com/view/ccfm-neur… Call for papers: sites.google.com/view/ccfm-neur… Submission Link: openreview.net/group?id=NeurI…

Excited to share our new work: “Language Models Improve When Pretraining Data Matches Target Tasks” Yes, it sounds obvious (and it is!), but typically this only happens implicitly and indirectly: intuitively select data → benchmark → refine → repeat. We wondered: what happens if we explicitly match pretraining data to benchmarks? The result is a dead simple approach that yields 2x+ compute multipliers over strong baselines and gives us a principled way to study how benchmark choices shape (and constrain!) model capabilities. Bonus: extensive scaling laws from training 500+ models that reveal how optimal data selection evolves as models scale. 🧵 (1/14)

Terminal-Bench and @warpdotdev @zachlloydtweets in TechCrunch today :) (link in replies) I firmly believe that the future of LLM-Computer interaction is through something that looks like a terminal interface. Great to see this picking up steam.

Finally, if you are interested in continual learning for foundation models, we're also organizing a workshop at NeurIPS 2025! x.com/FartashFg/stat…

Please see our paper and code (github.com/apple/ml-tic-lm) for many more details and results. If you're also attending ACL, would also love to chat in Vienna! Very grateful to collaborate with these amazing folks during my #Apple internship: @Reza_armandpour @i_mirzadeh @sacmehtauw @Vaishaal @raviteja_vemu @OncelTuzel @MFarajtabar @HPouransari @FartashFg

Excited to share TiC-LM (Oral at #ACL2025)! LLMs can become outdated ⏲️ and re-training from scratch is costly💰. Ideally, we'd keep reusing and updating models on newer data ♻️. We study continual training as 114 CC months are revealed one-at-a-time. arxiv.org/abs/2504.02107

Today, I’m launching a deeply personal project. I’m betting $100M that we can help computer scientists create more upside impact for humanity. Built for and by researchers, including @JeffDean & @jpineau1 on the board, @LaudeInstitute catalyzes research with real-world impact.