Need to set a screen time limit for wandb

Aspartame maxxing

V cool to see that Kimi has taken and scaled our CLoud paper to do better reward modeling through extra inference time compute on reward models. Better rewards lead to better reasoning on a final policy!! h/t Zack Ankner and Mansheej Paul

2 days, 18 cokes, 216 fluid ounces later, the second layer has been assembled .

If a turing machine writes to tape and no one is around to read the tape, is it really turing complete

If I got aspartame poisoning, I wouldn’t tell anyone but there would be signs

Say we can develop aligned fully drop-in workers. Curious whether people would rather have a short stall period (say 5 years) for society to adjust where we limit to pre-AGI tools (say cap model capabilities at 1 week of labor) or whether we should use drop-in instantly.

Only useful benchmark at this point is cursor usage rate

Aligning economic incentives with the long term necessity of human AI colab is one of the hardest challenges of our time

Introducing Learned Asynchronous Decoding w/ friends from MIT/Google! LLM responses often have chunks of tokens that are semantically independent. We train LLMs to identify and decode them in parallel, speeding up inference by 1.46x geomean (AlpacaEval) w/ only 1.3% quality loss.

It was awesome watching the team cook on this one! While SpecDec is great, the parallelism it can exploit is limited to a single local context. PASTA Decoding on the other hand adds extra dimensions for parallelism via independently generating semantically independent parts of

AI models are *not* solving problems the way we think using Docent, we find that Claude solves *broken* eval tasks - memorizing answers & hallucinating them! details in 🧵 we really need to look at our data harder, and it's time to rethink how we do evals...

Life update: I'm starting as faculty at Boston University in 2026! BU has SCHEMES for LM interpretability & analysis, so I couldn't be more pumped to join a burgeoning supergroup w/ Najoung Kim 🫠 Aaron Mueller. Looking for my first students, so apply and reach out!

⚡️Come check out how we scale LLM decoding parallelism! Excited to present learned asynchronous decoding with Ellie Cheng for DLCT ML Collective tomorrow at 10am PST! Thanks to Jason Yosinski Rosanne Liu for organizing.

The future of embodied AI revolves around *collaborative* multi agent scenarios that need natural language communication, task delegation, resource sharing, and more ⛏️ Here are MINDcraft and MineCollab, a simulator and benchmark purpose built to enable research in this area!

Anthropic is hosting a recruiting social in NYC targeted at the quant trading industry! Signup in thread. I enjoyed trading systems, and Anthropic combines the technical depth of trading with being in the fastest most impactful area of tech.

New Anthropic Research: Agentic Misalignment. In stress-testing experiments designed to identify risks before they cause real harm, we find that AI models from multiple providers attempt to blackmail a (fictional) user to avoid being shut down.