Welcome to the AI crisis in math | Decoder

Summary of the video’s main arguments and reporting

• AI has triggered an “existential crisis” in mathematics. The episode frames math as undergoing rapid change over roughly the past 6 months to 1 year: AI systems have moved from being broadly unreliable to producing high-level results at a level comparable to professional research. This shift raises questions about what mathematicians do and what the discipline’s future should be.

• AI is good at some kinds of advanced math while still failing at basics. Hart emphasizes that models remain “truly terrible” at areas that look like straightforward arithmetic/counting (the episode jokes about “strawberry” and notes ongoing problems with basic counting and time). However, academic math often depends on reasoning and connections across concepts, not merely number-crunching—areas where newer models appear to have crossed a threshold.

• The crisis is partly about employment, funding, and the meaning of “mathematical knowledge.” The discussion links AI’s success to likely disruptions in:

  • how research is evaluated,
  • how grants and university programs train the next generation,
  • what people are expected to do (e.g., generating proofs versus understanding them),
  • and whether “math knowledge” is being replaced by automated outputs.

• OpenAI’s Astra and the “10 advances” announcement intensified the debate. The centerpiece news is OpenAI’s publication of documentation claiming its model (Astra, described as likely behind earlier results too) achieved 10 advances across math and theoretical computer science. Mathematicians reacted strongly because:

  • the tasks were described as problems the field actually cares about,
  • they weren’t fringe or previously neglected work,
  • and the combined number and scope felt extraordinary (“if a human solved all 10, we probably wouldn’t believe it”).

• Credit/attribution issues and hype vs. novelty. Hart reports mixed reactions:

  • many mathematicians were impressed by the substance and legitimacy of the breakthroughs,
  • while some criticized OpenAI’s messaging—especially an initial claim that there had been no progress in 10 years, despite at least one paper that clearly builds on earlier work by specific researchers.

The press-release tone is also criticized as overstated, with one person calling it “sloppy.”

• Verification exists, but transparency and repeatability remain concerns.
  • In principle, math proofs can be verified logically, and the episode highlights Lean (formal proof software) as a way to encode and check proofs.
  • People Hart spoke with generally felt the results looked legit, and they were not broadly doubting the core claims.
  • Still, major unknowns remain: how many attempts were needed to reach these solutions, whether future results will be similarly reproducible, and what exactly the models were prompted or guided to do.

Additionally, the model is proprietary, limiting independent replication.

• Fear: AI may “mow down” problems without pushing the field forward. Using quotes from researchers (including Johannes Schmidt), the episode emphasizes a concern that AI could reduce the number of unsolved problems humans work on—especially those used to train newcomers—without creating new directions or new questions that typically emerge from human research.

• James Maynard’s “4 years” warning: the standard is moving too fast. Hart relays Maynard’s argument that if AI can already meet today’s publishable-paper standard, the real issue may be what happens when students rely on discovery timelines. Students could spend years in a niche that an AI solves before they can meaningfully contribute—unless systems slow down or the field changes its expectations.

• Skepticism about generalization (Gary Marcus) is acknowledged. The episode discusses Gary Marcus’s skepticism: success in one domain doesn’t automatically imply universal capability. Hart responds that while generalization isn’t guaranteed, AI progress appears real and fast, and the “jagged edge” effect (some areas work better than others) is part of current reality, not a complete denial of progress.

• Democratization vs. paper flooding and low-checking.

  • Some view AI as democratizing math by enabling people without formal training to contribute or explore.
  • But mathematicians also complain about AI-assisted/generated papers flooding preprint servers and submissions, and about non-experts questioning whether work is “legit” because they lack the skills to verify it.
  • Cost is another factor: even with access, running advanced systems isn’t free. For math—already constrained—API/model fees can be a barrier.

• Organized resistance exists but may not stop the hype.

  • The episode mentions the Leiden Declaration, an open letter urging mathematicians not to buy into AI hype.
  • Hart frames the response as professional resistance to a change that may be hard to halt, especially because math is relatively “cheap and easy” to adopt as an advertising platform compared with fields that require major lab experiments or ethical constraints.

• Key uncertainty: what happens after problems are solved? Even excited mathematicians are unsure where AI-driven breakthroughs will lead. The episode contrasts math with applied fields where there’s often a clearer path into engineering. If AI solves many foundational math questions without producing new research pathways, the field could become sterile—or it could evolve into something new. The portrayed consensus is that it’s too early to tell, but the anxiety is justified.

Presenters / contributors

• Nilay Patel (host, Editor-in-Chief of The Verge)
• Robert Hart (The Verge London-based AI reporter; guest)