Caltech's AI solves 60-year math problem, hints at crisis prediction
AI trained on unsolved math puzzle can predict stock crashes years ahead.
A team of researchers at Caltech and other institutions published a preprint detailing how they applied reinforcement learning to the Andrews-Curtis conjecture—an unsolved 1965 math problem from group theory. The conjecture posits that any complex mathematical configuration can be reduced to a basic 'home' form using a finite sequence of three moves. For decades, mathematicians proposed counterexamples that seemed to disprove the conjecture, but no one could find the path to confirm or refute them. The AI navigated a game-like space with up to billions of steps, using only those three moves, and successfully found complete or partial paths for several unresolved counterexamples, showing they do not disprove the conjecture.
The achievement demonstrates that AI can operate in incredibly high-dimensional, long-sequence spaces far beyond human intuition. The senior author, Sergei Gukov, explains that this problem forced the team to develop AI systems that can adapt to extreme complexity. While the immediate result advances mathematics, the underlying method—efficiently searching vast possibility spaces with limited moves—could be applied to real-world data. The researchers suggest that similar AI may one day analyze massive datasets in finance, medicine, and climate science to detect subtle signals, enabling predictions of stock crashes, disease onset, or weather disasters years in advance, giving humans time to prepare.
- Caltech AI solved multiple counterexamples of the Andrews-Curtis conjecture that mathematicians couldn't crack for 60 years.
- Reinforcement learning AI navigated a maze with billions of possible steps using only three allowed moves.
- Method could scale to forecast complex events by finding patterns in massive, high-dimensional datasets.
Why It Matters
Turns unsolved math into a training ground for AI that may forecast global crises years in advance.