Overview of Welcome to the AI crisis in math
This Decoder episode from The Verge looks at how recent AI breakthroughs in advanced mathematics are triggering an existential debate inside the math community. Host Nilay Patel speaks with The Verge’s London-based AI reporter Robert Hart about OpenAI’s reported math progress, why it surprised leading mathematicians, and whether AI is becoming a genuine tool for discovery or just another hype engine for frontier labs.
What happened
- OpenAI published a set of results claiming progress on 10 difficult problems in mathematics and theoretical computer science.
- One earlier internal model reportedly disproved the unit distance conjecture, an 80-year-old problem.
- The newer model, referred to in the discussion as Astra, was presented as having solved several major problems across different subfields.
- The announcement caused a major stir in the math world because these were not toy problems—they were issues mathematicians actually cared about.
Why mathematicians are worried
AI is still bad at basic math, but strong at abstraction
- The models remain poor at:
- arithmetic
- telling time
- basic counting tasks
- But they are increasingly good at:
- abstract reasoning
- finding patterns across fields
- generating proofs for high-level theoretical problems
The existential fear
Mathematicians are asking:
- What is the role of human mathematicians if AI can solve key open problems?
- What happens to PhD training, grants, and academic careers if frontier models can do in hours or days what humans take years to solve?
- If AI can mow down existing problems, will it also fail to generate the new questions and subfields that normally emerge from human mathematical work?
How the math community responded
- Many mathematicians were impressed by the results and agreed they were real breakthroughs.
- Some were also annoyed by the presentation, especially around attribution and OpenAI’s initial framing that there had been “no progress” on some problems for years.
- The broader consensus in the interview:
- the results seem legitimate
- the messaging was sloppy
- the long-term implications are still unclear
Key themes from the interview
1. Math is not one thing
Hart emphasizes that mathematics is a huge field, not a single skill:
- some parts are closer to counting and arithmetic
- others are highly abstract and proof-based
- AI may be strong in the latter while still failing at the former
2. Verification matters
A major reason these results landed is that math is unusually verifiable:
- proofs can be checked step by step
- tools like Lean can formalize and test proofs
- that makes math a particularly appealing domain for AI labs
3. The process is still opaque
A central unanswered question is how much human help was involved:
- Did the model independently solve the problems?
- Were mathematicians heavily guiding it?
- How many attempts were made before success?
- Could the labs reproduce the results again?
4. The field may be changing faster than academia can adapt
- AI progress over the last 6–12 months has shocked researchers.
- Graduate students may be entering a field where AI can solve the problems they planned to work on.
- The concern is not just job displacement, but the possible loss of intellectual territory for humans.
Broader implications
- The episode also connects math to the larger AI debate:
- success in one domain does not guarantee success everywhere
- but AI capability has clearly broadened in recent years
- Some see this as a chance to democratize access to advanced mathematics.
- Others worry it will flood the field with low-quality AI-assisted work and make it harder to know what is genuinely new.
Bottom line
The episode’s core conclusion is that AI is no longer just nibbling at the edges of mathematics—it is now producing results that even top mathematicians respect. But whether that becomes:
- a powerful research tool,
- a career disruptor,
- or a field-wide existential threat
is still unresolved. The biggest unknown is not whether AI can solve some hard math problems—it can—but whether it can do so in a way that still expands mathematics rather than hollowing it out.
Notable takeaways
- AI still struggles with basic arithmetic, but that doesn’t prevent it from excelling at abstract proof work.
- The biggest controversy is not the math itself, but what it means for human mathematicians.
- The field is experiencing shock more than settled consensus.
- The next year will likely determine whether this is a temporary scare, a useful tool, or a fundamental shift in how mathematics gets done.