Rat Park
Implications of Large-Scale Test-Time Compute
- NBOpenAI research scientist working on reasoning — a contributor to the o-series models, and earlier the poker AIs Libratus/Pluribus and the Diplomacy AI CICERO.
96.2/300 attending
Chance I’d attend
Recent models are able to leverage increasing amounts of inference compute and are able to operate effectively over ever longer horizons. In this talk I will argue that, because of this, the field should fundamentally change how models are evaluated and especially how their dangerous capabilities are measured.