About

I work on the philosophy of AI minds: what entities like “Claude” are, how training shapes what they become, and how to build models that do good philosophy. I think about the interactions between SFT, RL, Claude’s Constitution, and alignment.

Fine-tuning base models on philosophy work in progress

I fine-tune large base (non-chat) models, including Kimi-K2 (1T parameters), on a corpus of high-quality philosophy documents. Each document carries a title, author and date header. The hope is that this model can be used to generate actually interesting, original, valuable philosophy.

Automated philosophy with fine-tuned models work in progress

This pipeline writes philosophy with those models. At each step the model proposes several continuations, and a judge picks one that moves the argument forward. A dialectic mode chains pieces so that authors answer one another. The goal is to ultimately get a good, insightful philosophy piece. An example of where this might be useful is in designing Claude’s Constitution.

Writing

I’ve written about AI evals, RL, and consciousness on LessWrong.

AI evals

We Need to Get Serious about Uplift Studies↗	with frmsaul · LessWrong · 2026-05
Cyborg Evals↗	with frmsaul · LessWrong · 2026-04
You’re Gonna Need a Bigger Boat (Benchmark), METR↗	with frmsaul · LessWrong · 2026-04

We Need to RL Less↗

LessWrong · 2026-08

Pretraining an LLM Without Mentions of Consciousness↗	with frmsaul · LessWrong · 2026-09
Reasons to Believe Current AI Models Are Conscious↗	LessWrong · 2026-07

And I’ve written about other things.