Mathematician Terence Tao posted a thread on Mathstodon on September 9, describing something the field is losing now that AI tools have entered mathematical research: finding a genuinely weighty open problem is itself becoming rare.
His claim is that good open problems — the kind that spawn follow-on work — are being mined in a non-renewable way. He added a sharper point: once word gets out that someone is working on a given problem, it can draw a wave of large-scale AI compute that races to solve it first, exhausting that research line before it has fully unfolded. In his words, this reverses centuries of open-science tradition.
The Difficulty Gradient Is Being Flattened
Tao's technical term for this is the difficulty gradient. In a healthy research field, problems are arrayed by difficulty: there are entry-level exercises, PhD-thesis-level problems, ones that require new tools, and a set of long-standing hard problems. Researchers use that gradient to decide where to go next, and to guide students.
In many fields, AI tools have flattened that gradient. A problem that once took three months and one that once took three years might now both get solved in the same afternoon — or neither gets solved at all, with no discernible line in between. He points out that there is currently no clear frontier separating "AI-solvable" from "AI-hard" problems.
This is amplified by another factor: AI companies generally don't publish negative results or disclose the process behind a solve once they get one. Outsiders see only the successes, not which problems were tried and failed. Tao's post does not name any specific company, nor does it give a concrete list of problems that have been "mined" — both points are currently unverifiable. Without failure data, the field can't even judge which kinds of problems are still safe to work on.
From Open to Closed
A knock-on effect he mentions is already observable: some mathematicians have started keeping their core research directions private.
Academia's default for the past few centuries has been to share ideas as early as possible — presenting half-finished work at seminars, posting unfinished lines of thought on preprint servers, writing up conjectures on blogs. That habit rested on the assumption that even if others found out, they probably couldn't finish it quickly, leaving you time to complete it yourself. Compute has now broken that assumption.
Tao's suggestion is to shift the evaluation standard one notch earlier: for certain problems, it shouldn't be enough to just supply an answer — the solving process and the associated difficulty structure should be analyzed too. Making "why this problem is hard" a publishable result in its own right is a way of rebuilding the gradient that has been flattened.
How This Connects to His August Paper
In August, Tao turned a public lecture from the International Congress of Mathematicians into a twelve-page arXiv paper on the theme of "proof glut": assuming AI can already do research-level mathematics, the field would face a situation where output outpaces anyone's ability to read it.
This Mathstodon thread looks at the other end of the pipe. The earlier paper worried about the output side; this one worries about the input side — where problems come from in the first place. Taken together, he's describing trouble at both ends of the same pipeline: good problems being consumed upstream faster than they can regenerate, while proofs pile up downstream faster than anyone can verify them.
Implications for China
China's math community may feel this mechanism more directly. First, the share of domestic teams discussing research directions on open preprint servers and social platforms has been rising in recent years, with many groups treating arXiv as their primary release channel. Second, the institutions able to mobilize large-scale inference compute to sweep open problems are concentrated in a handful of companies; on the Chinese side, most researchers are consumers of that compute rather than the ones directing it.
At a practical level, graduate student topic selection is likely to be affected. If a problem a PhD student has spent three years on gets solved by some company's model in a week during year two, how those three years should count has no corresponding clause in any university's degree-evaluation system. Over the past year there have already been several authorship and priority disputes, and there is no precedent yet for how to resolve this kind of conflict.
Tao himself offers no institutional fix; he leaves it at saying the field needs to rethink what it rewards. Whether that statement translates into actual changes to evaluation rules should become visible within a year or two: if journals start accepting papers whose main contribution is difficulty analysis, that path will have worked; if not, the retreat from open discussion will continue.
Sources: Terence Tao's Mathstodon posts, arXiv, CocoLoop, Simon Willison's blog; the phrases "non-renewably mined" and mathematicians no longer disclosing research directions are rendered per Tao's original English wording, and the section on implications for China is editorial extrapolation.