The Human Substitution Test as a Sanity Check for AI Evaluations
TL;DR: We suggest a sanity check for proposed evaluation or AI oversight schemes: Imagine the AI was replaced by a competent, strategic human — someone who…
Latest
Cap-and-trade question: AI-2040
TL; DR: I find it hard to believe that сap-and-trade from AI-2040 would be effective in a context involving prohibitions. This will likely create an unsustainable…
Plan A's problem with dry tinder
A group is worried about an approaching fire spreading rapidly through their city. They manage to halt the fire outside the city gates. Meanwhile they build…
AI #176 Part 2: Plan B
This is part 2 of the weekly, broadly covering speculation, rhetoric and policy, along with alignment research. This does not cover the release of GPT-5.6-Sol. As…
Beliefs and position mid 2026
See 1 Jan 2026, 1 Jan 2025 and the July 2025 update. This continues my habit of documenting my beliefs and feelings as we transition to…
Stories of the future are undermined by agent assumptions
Part 1 of a series on process alignment, a different way of thinking about agents at every scale.IntroductionThere are quite a lot of disagreements in the…
Value generalisation: value correction
I firmly believe that value generalisation[1]is the key to AI Alignment. That, indeed, it is necessary and almost sufficient for alignment.But I won't be arguing that…
Don't normalize a permanent underclass (even a rich one)
Epistemic status: my first solo post here, so critiques of both substance and form are welcome. Also, note that, while I wrote this and all of…
Reading into VLM hallucinations using the Jacobian lens
Reading and editing the visual workspace of a vision-language model.Vision-language models hallucinate: ask one whether some object is in a picture and it will happily say…
Are We Guarding Against Backdoors Or Failing To Notice Them? (Part 1 / 6)
This post serves to argue that backdooring evaluations are prone to failures stemming from triggers never reaching models.In backdooring literature, there is a common workflow. Outputs…