Princeton’s Kothari says AI did most of a math advance, with humans steering

Princeton computer scientist Pravesh Kothari says his contribution to an AI-assisted mathematical advance was so limited that, a year earlier, it would not have earned him authorship. In comments published October 13 by Princeton Engineering, he estimated that the work took weeks rather than years—though humans still had to redirect agents that became stuck.

The collaboration brought together researchers at Princeton, UCLA and the University of Texas at Austin. Their target was the Grothendieck constant, whose precise value remains unknown. The team used AI to tighten the limits within which that value must fall, rather than determine it exactly.

That result predates Kothari’s new comments. The case study’s arXiv record lists an initial submission on August 11, 2026, and a revision on August 14. The paper detailing the mathematical results has not yet been peer reviewed.

Two agents, then a change of question

A reasoning model worked on the mathematics, while a coding agent checked its proposals. Human instructions went onto a bulletin the agents read periodically. The system maintained a permanent memory so successive rounds could build on earlier work.

Around iteration 18, the researchers noticed that the agents were looping. They reframed the question and redirected the work. After roughly 30 more iterations, the system succeeded in narrowing the bounds. Kothari said the models did most of the research, with humans supplying guidance when progress stalled.

If I had the same level of involvement on a paper last year, I would not be listed as an author

Pravesh Kothari, speaking to Princeton Engineering

Finding answers versus choosing directions

The authors’ assessment, reflecting models as of August 2026, distinguishes solving a problem from judging its importance. They found models weaker at valuing theoretical results over numerical evidence, judging how broadly results apply, and identifying their significance or later uses.

The researchers attribute that gap to training material. Papers, textbooks and solved problems rarely preserve the full discovery process, including failed approaches and changes in strategy.

Kothari also described a teaching consequence: some first-year graduate starter projects no longer serve their old role as practice for harder research. His optimistic possibility is that researchers recalibrate those steps. His pessimistic one is that people abandon big problems because they fear a model will beat them to the result.