One hundred years ago, a 23-year-old postdoc named Werner Heisenberg completed a calculation that would become the heart of quantum mechanics, a radical yet stunningly accurate theory of the atomic and subatomic world. quantamagazine.org/its-a-mess-a-b…

Are you afraid of LLMs teaching people how to build bioweapons? Have you tried just... not teaching LLMs about bioweapons? @AIEleuther and AI Security Institute joined forces to see what would happen, pretraining three 6.9B models for 500B tokens and producing 15 total models to study

This is a really thoughtful and grounded discussion of the tradeoffs in scientific comms!

Docent, our tool for analyzing complex AI behaviors, is now in public alpha! It helps scalably answer questions about agent behavior, like “is my model reward hacking” or “where does it violate instructions.” Today, anyone can get started with just a few lines of code!

I'm glad to see a serious LLM forecasting effort. These kinds of forecasts seem like an undersupplied public good. I think good policy-conditional long-term forecasts will play a big part in avoiding bad outcomes for humanity, if we can get them set up in time.

Last month we held a workshop on Post-AGI outcomes. Here’s a thread of all the talks! 🧵 x.com/DavidDuvenaud/…

MATS 9.0 applications are open! Launch your career in AI alignment, governance, and security with our 12-week research program. MATS provides field-leading research mentorship, funding, Berkeley & London offices, housing, and talks/workshops with AI experts.

I'll be a MATS mentor this winter! (Jan-Mar 2026) Come work with me on methods for improving dangerous capability evals, and understanding agent behaviours and goals. Apply by Oct 2nd – matsprogram.org/apply#Ududec

We at AI Security Institute recently did our first pre-deployment 𝗮𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 evaluation of Anthropic's Claude Sonnet 4.5! This was a first attempt – and we plan to work on this more! – but we still found some interesting results, and some learnings for next time 🧵

Several AI developers aim to build systems that match or surpass humans across most cognitive tasks. Today’s AI still falls short. Our new report maps progress and highlights the key barriers that remain🧵

I really like this research programme aiming to understand goal directedness from the bottom up! Great example of how to combine conceptual clarity with systematic experiments.