New Anthropic blog: We benchmark approaches to making classifiers more cost-effective by reusing activations from the model being queried. We find that using linear probes or retraining just a single layer of the model can push the cost-effectiveness frontier. 🧵1/

swyx Emmanuel Ameisen Anthropic Andon Labs All watched over by vending machines of loving grace

If that's not generalization than I don't know what is

Can endorse working with Jack, and can't think of a more fun subject. Come peer into the robot's mind!

Very excited for this work! We trained an agent to audit models, and evaluated it using the same auditing game that human evaluators participated in (including yours truly). It wins ~40% of the time! For context, some human teams never solved it. Others took hours.

Lots of good research came from the first iteration of the program, including an open source mech interp library to trace circuits (github.com/safety-researc…). Recommend applying if you're interested!

You're absolutely right.

In which the gang (Runjin Chen, Andy Arditi, Jack Lindsey ): - identifies vectors for bad personas (evil, sycophancy, hallucinations, etc) - shows that if you inject the bad vectors in training, the model learns to not do the bad thing!! aka vaccines but for LLMs

New research with coauthors at Paul Jankura, Google DeepMind, EleutherAI, and Decode Research! We expand on and open-source Anthropic’s foundational circuit-tracing work. Brief highlights in thread: (1/7)

Researchers from Goodfire, Google DeepMind, Decode, Eleuther, and Anthropic wrote a post about tracing circuits in language models! We cover how to train replacement models and compute graphs of model internals, and even filmed a 2-hour walkthrough of interpreting some examples!

The "Circuit Analysis Research Landscape" for August 2025 is out and is an interesting read on "the landscape of interpretability methods" and model biology Qwen3 4B is also out on Circuit Tracer

Our interpretability team is planning to mentor more fellows this cycle! Applications are due Aug 17.