marginalia

Things I worked out, written down before I forgot them.

2 notes · est. 2026
napkin math & half-proofs

most recent

Adversarial evals: what the benchmark doesn't measure is what the attacker exploits

A short argument for why capability benchmarks and adversarial red-teaming are measuring different things, and what to do about it.

"Rigorous limits, and a full life folded inside them."

— from the marginalia

1 note

The action principle in one line, then the derivation that shows why Euler-Lagrange is the only equation you need.