A Fields Medalist Who Wrote a Paper on AI Extinction Risk Just Joined OpenAI to Work on Safety
Number theorist Jacob Tsimerman is moving from the University of Toronto to OpenAI. His argument is that AI still runs on trial and error, and that mathematicians are the people who could change that.
Jacob Tsimerman, the number theorist who was awarded a Fields Medal this year, is joining OpenAI to work on AI safety. The Fields Medal is mathematics’ most prestigious prize, given every four years to a small number of researchers under 40, so this is a notable person to move from the University of Toronto into an AI lab.
His stated reason is worth quoting in substance. Tsimerman calls AI an extremely transformative technology and says society is not putting nearly enough effort into safety. His specific claim about what mathematicians bring is the interesting part: AI research today runs mostly on an empirical basis, meaning people try things, measure what happens, and keep what works, with very few guarantees about why these systems behave as they do. That is a normal way to do engineering and an uncomfortable way to do safety. Proving that a system cannot do something is a mathematician’s job, and almost nobody can currently prove much about a large model.
He is not a new voice on the risk side. Last year Tsimerman co-published a paper on “omnicide events”, scenarios in which AI contributes to human extinction. That framing is contested among researchers, and he seems aware of the trap, saying panic is not the right response but that the risks need honest assessment. He also expects AI to outperform humans at mathematical research before long.
The move lands in the middle of an ongoing argument about how much AI has actually achieved in mathematics. Former DeepMind CEO Demis Hassabis called the recent advances real progress but not yet a fundamental breakthrough, reaching for AlphaGo’s famous “Move 37” as the standard, the moment a system produced something genuinely outside human intuition rather than a very good version of what humans do. For mathematics, Hassabis suggests the equivalent bar would be cracking something like a Millennium Prize Problem, the set of seven famous unsolved problems. OpenAI’s forthcoming Astra model has not managed that either, though Hassabis says he sees no reason it could not happen eventually.
What this means for you: nothing changes in any tool you use, and there is no product here. What it is worth reading as is a signal about where the difficulty now sits. The AI labs are hiring pure mathematicians for safety work, which is an admission that the current approach, build it and test it, does not produce the kind of assurances anyone would accept in aviation or medicine. If you have wondered why AI companies keep publishing evaluation percentages rather than guarantees, this is the reason: percentages are what empirical testing can give you. Whether that changes is a genuinely open question, and hiring people like Tsimerman is one of the more serious attempts at it.
Sources
Source: https://www.wsj.com/tech/ai/move-37-ai-demis-hassabis-google-deepmind-alphago-ec832a41
Readers Rated AI Short Stories Higher Than Human Ones, Right Up Until They Were Told a Machine Wrote Them
In three experiments with more than 2,500 people, readers did no better than chance at spotting ChatGPT fiction. The AI stories scored higher on quality and immersion, and the label alone was enough to drag the scores down.
