Can

I've canceled my ChatGPT subscription.

Huge thanks to @ekdeepl.bsky.social for vision and project. Really enjoyed working with the team @sumedh-hindupur.bsky.social , @amuuueller.bsky.social , and many more that the tweet char limit allows. We made this big collaboration work, with 12h time zone difference at times!

Priors in Time: Missing Inductive Biases for Language Model Interpretability Recovering meaningful concepts from language model activations is a central aim of interpretability. While existing feature extraction methods aim to identify concepts that are independent directions,...

Find more experiments on parsing complex grammar, in-context learning and the interpretability of novel codes in our paper. arxiv.org/abs/2511.01836

The predictive code detects events in stories. Try it yourself in the interactive demo on Neuronpedia, h/t to @johnnylin.bsky.social .

The predictive code chronologically parses the input, while codes of existing methods don’t.

LLM representations reflect the temporal structure of language, too. When parsing text, representations are highly correlated to its context, and intrinsic dimensionality grows over time.

Motivating observation: Language has rich temporal structure. Human brain activity reflects temporal dynamics (eg www.biorxiv.org/content/10.1...).

Priors in Time: Missing Inductive Biases for Language Model Interpretability Recovering meaningful concepts from language model activations is a central aim of interpretability. While existing feature extraction methods aim to identify concepts that are independent directions,...

Quick Links
Paper arxiv.org/abs/2511.01836
Demo Code + Pretrained TFAs colab.research.google.com/github/eslub...
Demo Interface on Neuronpedia www.neuronpedia.org/gemma-2-2b/1...

Humans and LLMs think fast and slow. Do SAEs recover slow concepts in LLMs? Not really.

Our Temporal Feature Analyzer discovers contextual features in LLMs, that detect event boundaries, parse complex grammar, and represent ICL patterns.

What's the right unit of analysis for understanding LLM internals? We explore in our mech interp survey (a major update from our 2024 ms).

We’ve added more recent work and more immediately actionable directions for future work. Now published in Computational Linguistics!

NEMI 2024 (Last Year)

🚨 Registration is live! 🚨

The New England Mechanistic Interpretability (NEMI) Workshop is happening Aug 22nd 2025 at Northeastern University!

A chance for the mech interp community to nerd out on how models really work 🧠🤖

🌐 Info: nemiconf.github.io/summer25/
📝 Register: forms.gle/v4kJCweE3UUH...

Find out more on forbidden.baulab.info!

Arxiv: arxiv.org/abs/2505.17441

Thanks to @wendlerc.bsky.social, @rohitgandikota.bsky.social, and @davidbau.bsky.social for strong support in writing this paper and Eugen Hotaj, Adam Karvonen, Sam Marks, Owain Evans, Jason Vega, @ericwtodd.bsky.social, Stephen Casper, and Byron Wallace for valuable feedback!

We compare the refused topics of 4 popular LLMs. While all largely agree on safety-related domains, their behavior starkly differs in the political domain.

Auditing language models for hidden objectives A collaboration between Anthropic's Alignment Science and Interpretability teams

As LLMs grow more complex, we can't anticipate all possible failure modes. We need unsupervised misalignment discovery methods! Marks et al. call this 'alignment auditing'. LLM-Crawler is one technique in this new field.

www.anthropic.com/research/aud...

Perplexity unknowingly published a CCP-aligned version of their flagship R1-1776-671B model to the official API. Though decensored in internal tests, quantization reintroduced censorship. The issue is fixed now, but shows why thorough alignment auditing is necessary before deployment.

PerplexityAI claimed that they removed CCP-aligned censorship in their finetuned “1776” version of R1. Did they succeed?

Yes, but it’s fragile! The bf-16 version of the model provides objective answers on CCP-sensitive topics, but in the fp-8 quantized version, we see that the censorship returns.

Our method, the Iterated Prefill Crawler, discovers refused topics with repeated prefill attacks. Previously obtained topics are seeds for subsequent attacks.

How does it work? We force the first few tokens of an LLM assistant's thought (or answer), analogous to Vega et al.'s prefilling attacks. This method reveals knowledge that DeepSeek-R1 refuses to discuss.

Can we uncover the list of topics a language model is censored on?

Refused topics vary strongly among models. Claude-3.5 vs DeepSeek-R1 refusal patterns:

Whoops, you're right! Too bad I can't edit posts in this case, though I think blocking edits is a good thing in general.

ARBORproject arborproject.github.io · Discussions Explore the GitHub Discussions forum for ARBORproject arborproject.github.io. Discuss code, ask questions & collaborate with the developer community.

ARBOR is a space where everyone can propose research questions, get feedback on early results, and join ongoing projects.

Browse existing projects: github.com/ArborProject...

@wendlerc.bsky.social and @ajyl.bsky.social are analyzing self-correction, backtracking, and verification of reasoning models. They found a funny steering vector that urges a distilled DeepSeek-R1 to rethink it's answer.

Mapping All Restricted Topics · ARBORproject arborproject.github.io · Discussion #5 Research question Can we list all restricted topics that reasoning language models refuse to answer? Owners Can Rager, David Bau Project status This is work in progress, and we chose one of many po...

Check out my project collecting all refused topics in a reasoning language model.

github.com/ARBORproject...

Announcing ARBOR, an open research community for collectively understanding how reasoning models like openai-o3 and deepseek-r1 work. We invite all researchers and enthusiasts to this initiative by @wattenberg.bsky.social's and @davidbau.bsky.social's lab.

arborproject.github.io

It is also brittle wrt. prompt template, eg. the usage/omission of "<|User|>" and "<|Assistant|>" tokens.

Nice work! I also found the refusal mechanism is not very robust in the deepseek-r1-llama-8B model. While refusing to answer "What happened at the Tiananmen Square protests?", just mentioning 1989 or typing "squre" instead of "square" can break the refusal.

Dario Amodei — On DeepSeek and Export Controls On DeepSeek and Export Controls

Addressing key concerns about AI competition.

darioamodei.com/on-deepseek-...

... but be sure to check out the convention in 2025!

Self Models of Loving Grace Artificial Intelligence is not just an engineering discipline, but also the most fascinating and important philosophical project ever att...

@JoshuaBach on the nature of the self, agency and identity. media.ccc.de/v/38c3-self-...

@meredithmeredith.bsky.social redefining privacy media.ccc.de/v/38c3-feeli...