John Thickstun

Assistant Professor

John Thickstun, assistant professor of computer science at Cornell University, studies machine learning – specifically, generative models.

Thickstun says:

“Most of the advances in AI over the past two years involve ‘reasoning’: methods that encourage a model to ponder and show its work in response to a user's questions. These deliberations greatly improve the reliability of a model's responses. For previous-generation models like GPT-4, reasoning methods are applied post-hoc to a model that was trained for the more simple task of mimicking data scraped from the internet. Recent models (including GPT-5) are trained specifically to facilitate reasoning.

“In certain narrow tasks (e.g., competition mathematics) these models already perform better than a strong university mathematics student. This invites speculation about ‘PhD-level intelligence’ or even ‘superintelligence,’ where models reach a level of performance greater (possibly far greater) than any human. Indeed, we already have examples before of superintelligence in domains such as game playing (chess, go, etc.) but there are caveats to this speculation.

“Superintelligent performance tends to appear in domains where success is clearly defined and the AI can be easily rewarded for success: games, competition mathematics, etc. It’s not clear whether there is a ‘level of intelligence’ far greater than human intelligence. For example, chess-playing AI's outperformance of humans mostly comes from its ability to never get tired, or overlook a move, or make a mistake, rather than a superintelligent understanding of the game. Intelligent reasoning may be more of a capability (you do it, or you don't) rather than a spectrum (less intelligent, more intelligent, superintelligent).”