Now

What I'm working on this month. Updated May 6, 2026.

Work

Doing the BlueDot Technical AI Safety course, facilitated by BAISH (Buenos Aires AI Safety Hub).

Reading

El infinito en un junco by Irene Vallejo.

Writing

Drafting a paper on which instructions LLMs actually retain across long coding sessions (Not All Instructions Are Forgotten Equal). Bayesian ordered logistic over 244 compliance observations; treatment effects span an order of magnitude across instruction types. Read the draft (PDF).

Thinking about

AI safety, particularly how you tell whether a system has internalized a rule versus pattern-matched around it.


This is a /now page, a snapshot of my current focus. Older snapshots live on /then.