Then
A record of past /now pages.
May 18, 2026
Work
Doing the BlueDot Technical AI Safety course (facilitated by BAISH, the Buenos Aires AI Safety Hub) and self-studying ARENA on the side.
Reading
El infinito en un junco by Irene Vallejo.
Writing
I wanted to know why LLMs follow some instructions all session but forget others within a few turns. So I ran an experiment: 244 cases of an instruction being obeyed or ignored, fit with a Bayesian ordered logistic model. Some instruction types got followed about 10x more reliably than others. Read the draft (PDF).
Thinking about
AI safety, effective altruism, and how to convince people to live a cognitively active life. If you want help disconnecting from social media and building more robust cognitive habits, chat with me!