Thinking Machines’ Safety Research Grants highlight tamper-resistant safety training, citing my work on Tamper-Resistant Safeguards for Open-Weight LLMs.
Locus, our autonomous AI research system, plans and runs experiments over multiple days to improve language models after their initial training. On PostTrainBench+, a benchmark for this post-training process, it achieves 51.6%, exceeding the official human-tuned Qwen3-1.7B-Instruct model (49.4%). Also, an LLM fully post-trained by Locus is currently running in enterprise production.
NanoGPT-Bench tests whether AI coding agents can discover ways to train GPT-2 faster, using a fixed computing budget. Given 512 hours of H100 GPU compute, existing SWE coding agents reproduce just 9.3% of the training-speed improvement achieved by human researchers over five months.
We introduce Locus, our autonomous AI research system, which plans, runs, and learns from experiments over multiple days. In a continuous 64-hour run on RE-Bench, a suite of AI research tasks, it scores 1.30 against the human-expert baseline of 1.27. It also earns medals in 77% of the machine-learning competitions in MLE-Bench Lite.
I'll be joining Intology to work on scaling autonomous AI research.
Our work on Tamper-Resistant Safeguards is featured in Wired. article.
I'll be joining the Center for AI Safety as a Research Engineer Intern.
WMDP is featured in TIME. article.
Announcing our $6M seed round and the launch of Mindy. article.
I gave a talk internally at Google Research. post.
