joining the culture war on the side of the Culture

Why is interpretability the key to dominance in AI? Not winning the scaling race, or banning China. Our answer to OSTP/NSF, w/ Goodfire's Tom McGrath Transluce's Sarah Schwettmann MIT's Dylan HadfieldMenell resilience.baulab.info/docs/AI_Action… Here's why:🧵 ↘️

Quality metrics are key to understanding and using SAEs, and to-date have been sadly lacking. We at Decode Research are happy to have been able to help support this project and are excited that 1.0 has been released!

We at Decode Research are very excited to make this available!

This is excellent work. My own timelines aren't quite this short, but as the authors suggest, this outcome is plausible enough to take into consideration ("plan for the worst"). It's realistic, detailed, and makes the AI takeoff feel much more plausible.

gemini 2.5 pro max is amazing at coding, but I wish it didn't have an ocd-level of obsession with fixing line length and linter errors (regardless of what I say in my rules file)

A surprisingly very positive UBI result. I hope we see more replications of this kind of experiment.

🚀 New stuff on Neuronpedia - 🎙️ Podcast: The Babble by NotebookLM - 🔍 TopK Search by Token: a new way to search + API - 🪆 Matryoshka SAEs by David Chanin - 🔬 Probity, a probing library by Curt Tigges - 🧠 New Auto-Interp Models Demos +examples in thread➡️

An excellent essay, and exactly why we're focused on bringing open-source interp tools to as many researchers as possible.

Danielle Fong 🔆 Reinforcement Learning with Wormtongue Feedback

yes, gemini, I know I was right to ask that question. that's why I asked it

Dear MAGA friends, I have been worrying about STEM in the US a lot, because right now the Senate is writing new laws that cut 75% of the STEM budget in the US. Sorry for the long post, but the issue is really important, and I want to share what I know about it. The entire

claude's recipes are a little unhinged, but I'm intrigued