The Turichron

Evening brief, Oct 26

AISI tests two limits of biological-model safeguards

In one study, fine-tuning restored filtered ESM-2 capabilities for about $80 of compute at the largest tested scale; separately, ReproBAIT agents rebuilt selected simpler-tool functions with supplied training data, but failed to reproduce RFdiffusion3.

AISI's filtering study · AISI's ReproBAIT study

Read the full brief