Maths is Cooked: AI's Latest Breakthrough -- And What's Next

Overview

The video argues that AI may soon make large parts of “frontier” mathematics effectively obsolete—at least in terms of finding solutions. It simultaneously warns against overhyping and predicts a more nuanced future: mathematicians will increasingly use AI rather than being fully replaced.

Main Points and Claims

• “Math is cooked” claim (rapid progress): The video presents physicist/computer scientist Alexander Wissner-Gross as arguing that AI is on pace to solve essentially all top-tier, professional research-grade mathematics within roughly 4–5 years, implying a near-term collapse of traditional mathematical bottlenecks.

• Acceleration in AI math ability: The presenter claims large language models have improved dramatically—shifting from basic arithmetic failures to solving high-level problems, including some “no one had heard of before.”

• Headline event: an AI disproved a longstanding conjecture: The video highlights an event described as happening “two weeks ago,” in which an internal OpenAI model allegedly:

  • Disproved an ~80-year-old conjecture by Paul Erdős.
  • Produced a counterexample that refuted the conjecture (Erdős had conjectured that a square lattice form was optimal; the AI found a counterexample).
  • Did so on the first attempt (“one-shotted”).
  • Achieved correct answers in repeated tests with enough compute (subtitles suggest a rate “%,” but the exact figure is unclear).

• Caveat: reliability and “logic checking” still weak: Even with the headline-worthy counterexample, the video notes the model is not great at verifying its own reasoning. It also emphasizes that the solution’s significance depends partly on it being understandable/real, not just producing “mathy” text.

• Mathematicians’ emotional/political reactions: Subtitles include quotes reflecting existential anxiety—concern that AI could outperform human experts and undermine academic careers. Scott Aaronson is cited describing quickly hearing the news and sensing a gloomy mood among young researchers.

• Interpretive argument: taste and methodology vs intelligence: The video claims AI may not be “smarter” in a human sense, but may have worse “taste”—tending to produce unintuitive constructions humans would not naturally pursue. It also says the human community improved the AI result within hours, suggesting the “demise of math” narrative may be exaggerated.

• Pushback from the mathematical community: A group of 150 mathematicians is said to have signed a declaration warning that these developments threaten the autonomy of mathematics. The video argues that technology companies and media outlets have incentives to overstate AI capabilities.

• Prediction: mathematics won’t end; it will change: The presenter’s view is that AI will not fully eliminate mathematics, but will shift mathematicians toward AI-supported workflows. The argument is that much existing literature already contains enough information for AI-assisted proof attempts and fast search across many candidate arguments.

• Open question: can AI invent new methods? The video distinguishes between:

  • Likely scenarios: AI extracting patterns, filling gaps, and recombining existing literature.
  • Uncertain scenarios: AI developing fundamentally new proof strategies or methods.

• Skeptical stance on large language models: The presenter repeatedly argues that large language models alone are unlikely to reach the stage of truly inventing new mathematical methods independently.

• Overall conclusion: The video’s framing is “Math maybe cooked, but only partially”—more like “al dente” than fully cooked. The expected outcome is disruption, not a complete end of mathematics.

Sponsored Segment

The subtitles include a promotion for Incogni, a service that helps automate opting out of/removing personal data from data brokers, with a discount for early users.

Presenters / Contributors Mentioned

• Alexander Wissner-Gross (physicist and computer scientist; commentary/quotes)
• Scott Aaronson (computer scientist; cited quote)
• Paul Erdős (historical mathematician; subject of the conjecture)
• OpenAI (associated with the reported internal model result)
• Incogni (sponsor mentioned in the video)