Ben Recht

Machine learning research should shift toward design-based approaches and develop better theoretical explanations for empirical phenomena in model scaling and competitive testing.

“Author of "Prompts for Open Problems"”

The foundational concept of a data-generating distribution in machine learning theory is a myth; all randomness in ML is created or imagined by engineers rather than inherent to nature.

“Author of "There is no data-generating distribution" in Argmin”

Reinforcement learning as a computational paradigm is brutally inefficient and rarely produces desirable results in practice, despite recent success in fine-tuning language models.

“Author of "There's Got to be a Better Way!" in Argmin”

Reinforcement learning is fundamentally an iterative optimization process where a computer program receives external feedback scores and updates its code to maximize average performance, rooted in psychological theories of human learning.

“Author of "Defining Reinforcement Learning Down"”

The reinforcement learning community has fundamentally mischaracterized rewards as external environmental signals rather than internal agent computations, a mistake at the formalization level that can be corrected.

“Wrote about the Reinforcement Learning setup and how computer-science RL may have gotten rewards wrong.”