Bio
My work is often about bad epistemics and overoptimisation. I've done a few things. I run a research company, Arb. I got a PhD in AI from Bristol. I wrote a book with Dwarkesh Patel about how scaling happened. In the last pandemic, I worked on the effects of masking and government policy and advised the UK government.
Now president of Paradigm 3, a research nonprofit looking at capability generalization and open problems in pacing AI.
gleech.org
Projects
Research Direction: What are the biggest technical and strategic uncertainties in pacing?
The problem: most pacing proposals focus on one path through policy space where we will need a flexible response to new information and a changing technical and social situation.
The work: Substantially, we would answer these questions and also criticise others' public proposals for pacing.
The output: two papers, three or four case studies, two policy engagements.
Why it matters: The world will do some intervention and by default it's not going to be a good one. In my experience posting is very powerful because people really want good information.
Room to shape it: any of the 23 vague directions is great, other things we've missed also great.
What I'm looking for in a Mentee
Technical people who want to work on policy, policy people who want to pull the rope sideways by focussing on mechanisms and details. Anyone who's interested in these questions: https://pacing.tech/appendices.
What I'm Like as a Mentor
I'm mostly an ideas guy these days, but I don't insist on people working on my ideas if yours fit you better. I prefer text and respond same-day to emails. I care a lot about writing and will help finalise drafts with you.