Tomaso Poggio

Calibration can be effectively improved during training by focusing on curvature and margin dynamics, leading to better confidence estimates without sacrificing model performance.

Imagine slashing the human effort needed to go from hypothesis to submission-ready ML theory paper by orders of magnitude.

Momentum's impact on sharpness isn't straightforward: it either amplifies stochastic fluctuations to favor flatter regions or recovers classical stabilization to favor sharper regions, depending on batch size.

Identical prediction accuracy in financial time series can mask a 3x dispersion in portfolio turnover, revealing how optimizer choice acts as a hidden prior that dramatically alters model behavior.

Tool design, not just model size, is the bottleneck for LLMs to achieve "superintelligence" via the Diligent Learner framework.