Mathematician argues AI can 'strip mine' well-developed areas of math

On October 7, blogger onionesque posted a thread proposing a framework for judging whether AI can crack a math problem: if a problem has been attempted many times and its neighborhood is well formalized, AI has a chance to "grind through" it; historically, many problems were merely proxies, with the real value being the human effort they attracted toward building a field's conceptual infrastructure.

Confirmed

• Core metaphor: math is a "strip-mined" rich ore — the expensive "geological surveying" (mapping a field's terrain, inventing the right language and definitions, identifying promising directions, intermediate lemmas, etc.) was long ago done by humans, and AI can systematically mine these veins directly
• The traditional mainstream path of math research: notice an interesting phenomenon → pose a conjecture → the conjecture becomes a community rallying point → develop new tools (i.e., the real progress) → obtain answers. The author judges that as AI capability grows, this path relying on communities rallying around conjectures will soon no longer be mainstream
• In the short term, "de-beautifying" existing proofs or proof attempts is already useful, even if the original proof is ugly; in the medium term, AI may make it easier to build theories across fields, something that was hard in the past due to extreme specialization

Not yet confirmed

• For upstream conceptual development itself, the author says it's unclear what it would look like with AI alone
• The author admits curiosity about whether AI can attack problems that are important but relatively isolated
• He also notes the limitation: the applications above still only work within the boundaries of the "conceptual geology" mathematicians spent centuries building

Why it matters

The framework offers an actionable criterion for assessing AI's real capability in math research — not the prestige of a problem, but the maturity and verifiability of the conceptual infrastructure around it. Already-developed eps neighborhoods are both the ore and the path to verifiability, suggesting AI's near-term impact will concentrate in mature fields rather than opening up entirely new theoretical directions.

2026-10-07 ~ 2026-10-07 · 7 related posts

• [source] Where AI can grind through math: well-formalized neighborhoods, not isolated frontiers — _onionesque · 2026-10-07
• Math is full of 'strip-minable' deposits where the expensive conceptual groundwork is already done — _onionesque · 2026-10-07
• [source] The open question: can AI attack important but isolated math problems? — _onionesque · 2026-10-07
• The 'strip mining' research paradigm — phenomenon to conjecture to machinery — will soon matter less — _onionesque · 2026-10-07
• Math as a strip mine: why AI can exploit decades of pre-mapped intermediate results — _onionesque · 2026-10-07
• [source] The old route to math advances via conjecture-gathering will soon matter less, author argues — _onionesque · 2026-10-07
• AI proof helpers still trace the conceptual hull mathematicians built over centuries — _onionesque · 2026-10-07