Gwern's dictum resurfaces: sampling reveals presence of knowledge, never absence

akbirthko · x · 2026-10-07

akbirthko resurfaced Gwern's well-known dictum — "sampling can only reveal presence of knowledge, not absence" — noting that in hindsight, it should have been enough to explain a lot.

The point is a classic in LLM evaluation: a failed generation doesn't prove a model lacks some knowledge or capability; it may just be sampling luck. It's a caution against every "the model can't do X" claim — negative conclusions need heavier sampling or sharper probes, not a single roll of the dice.

Original post →

More from AGI Musings

AGI Musings channel →