Code for "Large Language Models Generate Harmful Responses Using a Distinct Mechanism, Shared Across Harm Types" is now available! It includes B-TAP, the interp method behind our work, for localizing and causally intervening on critical weights. Code + project page + paper 👇🏻

Since a reporter recently reached out to me about this, and inspired by @ChrisGPotts sharing his own comments publicly, I thought I'd share my perspective here as well. I find the reassessment of superposition genuinely interesting, especially given how influential the superposition hypothesis has been. In particular, the discussion of how conclusions drawn from toy models may not carry over to real networks with real data is important. Clearly more work is needed to support the broader claims. I'd particularly like to see systematic evaluations of how well the proposed neuron interpretations generalize across inputs and models. Superposition hasn't played a major role in my own interpretability research. It's one hypothesis in a much broader field, with many different approaches and research questions. As such, these findings don't fundamentally change the questions I'm working on, but they add to a broader view on how models store information. For me, an important goal of interpretability is to go beyond explanations. Can our understanding help us predict model behavior, improve training methods or make targeted causal interventions? Our B-TAP method, for example, uses interventions analogous to brain lesions to study mechanisms of particular behaviors, without requiring a complete neuron-level explanation. But most importantly, understanding individual neurons is not the same as understanding complex model behavior. Even if most neurons are interpretable, we still need to understand how they interact to produce behavior. This is especially true if we want to apply interpretability to safety, which includes very complicated behaviors over many tokens and turns such as reward hacking and deception. Trying to apply interpretability to these realistic scenarios is what we should focus on.

Given the rise in submissions, should @iclr_conf substantially lower the acceptance rate, to around 10%?

I'm excited to work with @OzAmram in the incoming @cbai_ai cohort. He's making quite a leap from particle physics to AI safety, and I'm sure he'll do awesome things. It's encouraging to see talented researchers from other fields recognizing the importance of AI safety and choosing to contribute to it.

Be careful when walking around the COLM venue at night. I was out alone and got approached by a group of people to discuss continual learning

It was a packed day, full of interesting conversations and engagement! Thank you to all the researchers, speakers, and panelists for making this happen. See you next year!

Aaand we're on! Exciting day ahead of us...

In this new work led by @sumedh_hrs, we show that while LLMs represent numbers along curved manifolds, the comparison algorithm can be described with linear directions.

Arrived in San Francisco for COLM 2026! 🌉 On Wednesday at 4:30pm, we’ll present our work on the connection between creativity and hallucinations in LLMs, led by @glazer_neta. Come by to discuss! On Friday: the Actionable Interpretability workshop! We have a packed program this year... @ActInterp

Check out our new work, to be presented at @COLM_conf! We find mechanistic evidence that creativity and hallucinations are tied together. Working on this made me realize that some of the same mechanisms that make LLMs useful may also create failures.

I don't want to waste my energy reading something you couldn't be bothered to write.

Come work with me! I’ll be mentoring in this fall’s @cbai_ai fellowship. If you’re interested in interpretability + AI safety, especially understanding the mechanisms behind sycophancy, deceptions, harmful generations and hallucinations, apply to my stream 👇

It’s been interesting watching PhD students and junior researchers start working with coding agents and tools. It’s clear it can dramatically increase velocity. On the other hand, it’s much easier for things to get far off track.

Announcing the keynote speakers for the Actionable Interpretability Workshop at COLM 2026! @ActInterp @boknilev @banburismus_ @ChrisGPotts @dhanya_sridhar What would you ask them?

Accepted to #emnlp ✅ 🎉 Read about the (non-) robustness of uncertainty probes 👇🏼

Never mind the ethics of AI use, we need to establish the etiquette. Inviting me to read LLM-generated text without telling me that's what it is, as though you'd written it yourself, is _rude_

1/ Text-to-video models generate beautiful videos, but they still struggle to follow complex prompts. Relations like left/right, towards/away, on top of, behind, or even multi-stage instructions are often generated incorrectly. In our new paper, accepted to SIGGRAPH Asia 2026, we introduce CVG, an inference-time guidance method that improves compositional video generation without retraining or architectural changes.