← Back to blog

Bootstrapping a weekly research memo

February 4, 2026 · 4 min read

I want a lightweight ritual for sharing what I learn each week about LLM safety and interpretability. The goal is not a full paper summary, but a short, high-signal memo that helps me retain ideas and gives collaborators a quick window into what I am exploring.

Working definition: one memo per week, three bullets of insight, one open question, and a link to a paper or experiment worth following.

Why this format

A short weekly memo forces me to separate “interesting” from “important.” It is also a test bed for a future automated agent: if I can keep the structure predictable, a tool can eventually draft the post from notes, and I can edit and publish quickly.

This week’s seed ideas

Open question

How can we reliably detect subliminal learning signals before they show up in deployment? If detection is costly, can we create small, cheap proxy tests that still catch the risk?

Next week I will try converting paper notes into the memo template automatically, then see how much manual editing is truly needed.