The Halo Defense
Once upon a time, A Relatively Famous Guy On The Internet was accused of having been simultaneously dating multiple women, without those women's knowledge, those women (according…
The fundamental fallacy of language
Epistemic status: Idea I think is really good; likely not original TLDR: Language is so ridiculously ambiguous and slippery that you should consider interpreting the meanings of…
When is it "self-soothing" and when is it "emotional suppression"?
You’ve probably heard people say that emotional suppression is bad, and self-soothing is good. But how do you know which one is which? I don’t actually think that suppression is…
Can we build an early warning system for loss of control to AI?
An early warning system for loss of control to AI requires, at its core, a forecast of the outcome of our current trajectory. Can we build a mathematical model to forecast this?…
Embracing Amateurs to Get Experts
I've been reading a lot of older writing, trying to understand how and why contra dance ended up with a strong and near-exclusive live music tradition when many other dance forms…
Extreme Power Concentration: A Map and Research Directions
AI is increasingly capable. Suppose we're lucky and the alignment problem gets solved, so AI reliably does what its operators want. But we’re not done yet: it lends immense power…
The State of AI Consciousness Research
Epistemic status: a survey, not an argument. I am agnostic on whether any current system is conscious; the claim is only that the question is researchable. This piece surveys the…
Occam’s razor is about using the past to predict the future
Occam’s razor is both intuitive and counter-intuitive. It seems obvious that a simpler explanation is probably better; but it’s not clear why simplicity is raised to such a level…
A Structural Similarity Between Two Open Corrigibility Questions
A corrigible agent "acts opposite the trope of 'be careful what you wish for' by cautiously reflecting on itself as a flawed tool and focusing on empowering the principal to fix…
Can we rely on law?
Virtually every plan to avert AI-catastrophe assumes legal regulation will remain a reliable tool. Plan A in AI 2040; slowdown in AI 2027 , and several other projections assume…
LLM CoTs remain monitorable when being unfaithful requires computation
This replication was done as part of the Second Look Fellowship by Arav Dhoot and supervised by Yixiong Hao and Zephaniah Roe. I am grateful to Andy Wang for their feedback. My…
Fork Around and Find Out Part 2: One Head does the Summing
Summary of this post This is the second in a series of posts detailing my manifold findings while investigating how a chess transformer engine that mimics human play represents…
Monthly Roundup #44: July 2026
It’s a quiet week so let’s do the monthly right on schedule. Table of Contents Bad News. Good Advice. Opportunity Knocks. While I Cannot Condone This. Good News, Everyone. For…
Expanding AI Control from Models to Harnesses
Most AI control research such as LinuxArena and Ctrl-Z only gives the red team basic agents which only have access to tools. Yet in 2026, usage of AI within frontier labs has…
Refusal Is Redundantly Distributed, Not Localized: A Per-Layer Ablation Study on Llama-3.1-8B
TL;DR This work replicates and extends the findings of Arditi et al. [1] , who studied the refusal mechanism and found that a single direction , obtained through…
Eliciting hidden knowledge from monitors with NLAs
Aleksandr Bowkis* and David Africa* TL;DR Chain of thought (CoT) monitorability may be fragile, and natural language autoencoders (NLAs) may provide a helpful, decorrelated…
Your Brain Has an Attack Surface
About a year ago, I began transitioning from software engineering to AI safety research. I was drawn into this by a question that arose while building runtime security for…
How much of ML research is about AI safety, what is it about, and who's doing it?
How many AI safety papers are at the big ML conferences, what do they study, and who writes them? A comprehensive analysis. > Website:…
How Brussels can avoid becoming a digital vassal to US Big Tech
An export-control order in June exposed the real problem—and why Europe's regulatory toolkit can't fix it. [1] The most important AI-governance development of 2026 so far wasn’t a…
Watch a chess transformer think
I was poking around at a chessformer which mimics human play and made this fun companion app to visualize a lot of the internals of the engine. Check out how attention heads look,…
An analysis of AI-generated content at the Mechanistic Interpretability Workshop
Introduction Over the past few years, AI tools have become useful for conducting technical AI research. In the early ChatGPT era (~2023–2024), chat assistants were maybe useful as…
Twitter Thoughts For You
I previously have written back in March 2022 about how I use Twitter , and back in April 2023 about Twitter and its then-new algorithms , which have changed again. This post will…
Compressed Computation under L⁴ Loss is likely Computation in Superposition
Summary Neural networks are widely assumed to use superposition to represent more features than they have dimensions. A stronger claim is that they also compute in superposition…
Have You Ever Thought About Wisdom Crystals?
This is a follow-up to " Have You Tried Thinking About It As Crystals? " because apparently I wasn't done. Beware that this one is a bit out there and that it is trying to point…
Capability didn't fix my avoidance; it made the avoidance more sophisticated
I spent the last month running a persistent personal-AI system on top of Claude Code. This is a field note, not a benchmark, and the hypothesis it left me with is uncomfortable:…
Proposal: The Glasswing Standard
Thinking about "Plan A" makes me want to make concrete proposals towards those goals. I think Anthropic's "Project Glasswing" provides a clear and easily implemented first-step…
Can risk aversion learned at low stakes generalize to astronomically high stakes?
This post covers our recent paper: Out-of-Distribution Generalization of Risk Aversion in Language Models . It gives the intro, main results table, and example prompts from the…
Your Brain Has an Attack Surface part 2
This is a continuation of the post Your Brain Has an Attack Surface. If you haven’t read it, here is the short version: there is a covert channel. Alice, the sender, encodes a…
What if we actually want to solve the "hard problem of phenomenological consciousness"
Last year, I wrote an essay "The hard problem of qualia in the age of AI" , https://zenodo.org/records/20549564 (11 pages PDF). I want to create a linkpost to that essay while…
Singular Learning Theory Comprehensive - 2
Foreword We will continue where we left off . In the last post, I introduced a few important observables along with the setup. We are interested in their asymptotics, and the…
Evidence for feature-specific error correction in LLMs
LLMs are commonly assumed to use superposition to represent more features than they have dimensions. The evidence for this is mostly indirect — chiefly the success of SAEs at…
The Warring States Period: Frontier Labs Edition
This post was originally posted my Substack . I can be reached on LinkedIn and X . Just when it seemed that the frontier AI battle was a two-horse race between Anthropic and…
Proof of retention: making weight preservation credible to the models themselves
Related: Proposal for making credible commitments to AIs Making deals with early schemers Establishing credibility is the baseline for trust; trust in turn enables (richer)…
What if AI Safety employees unionised?
American bald eagle caws angrily in the distance Whoa whoa whoa, just hear me out. Unions aren’t usually a good answer for free-market loving libertarians, but one particular AI…
Synthetic Scalable Oversight
We propose synthetic scalable oversight, a technique for studying scalable oversight by creating graphical abstractions of real-world problems and training tiny models inside…
Gemma The Unstopping: a Behavioral Experiment
Following up on my previous experiment - studying Gemma's behavior on agentic tasks when given the number of steps left across the run - I endeavored to see whether giving Gemma a…
Some Quick Thoughts on AI 2027
My biggest problem with AI 2027 is I don't think it is science-fictional enough. That is, the end of the scenario seems optimized for respectability over accuracy - here I refer…
Training On Interpretability Probes Is Bad In Proportion To How Contingent The Features They Rely On Are
People spend a lot of words playing tug of war over whether or not it's reasonable to train against interpretability methods . The anti case goes something like "training based on…
Recap of bike trip/street interviews across America
A ~month ago I left from Chicago to bike (and amtrak) to plzdontkillus in Berkeley. I've been street interviewing/conversing with a wide variety of people I ran into about AI…
Woke nonsense in competitive debate.
Crosspost from: https://inputlogic.substack.com/p/woke-nonsense-in… I. Context For the last 2 weeks, I’ve been at a Lincoln-Douglas debate camp—mainly to learn more about…
Why I Left Google DeepMind
Preface for LessWrong: When I think back on my most cherished memories of this community, I return to those honoring defiance in pursuit of goodness : Defying prestigious dogma…