A Thousand AI Monkeys With a Thousand Typewriters Will Give You OpenAI’s Erdős Result

The standard story about OpenAI’s Erdős result goes something like this. An AI model made a creative leap, connecting two distant mathematical domains in a way no human had previously imagined, and in doing so disproved a conjecture that had stood for nearly eighty years. It is a story about machine creativity, about AI developing genuine mathematical intuition, about a threshold being crossed.

There is a more honest story. And the more honest story is actually more interesting, more analytically useful, and more accurate about what AI can and cannot do than the creative insight framing OpenAI is promoting.

The more honest story is the thousand monkeys story.

What Actually Happened

The infinite monkey theorem holds that a monkey hitting keys randomly on a typewriter for an infinite amount of time will eventually produce the complete works of Shakespeare. The theorem is mathematically true but practically useless because there is no mechanism for identifying which random output is Shakespeare without a human reading every attempt. The monkeys produce Shakespeare in theory. They cannot produce it in practice because the verification problem is unsolvable at the scale the search requires.

Now change one thing. Give the monkeys a machine that tells them instantly, automatically, and unambiguously when they have typed a sonnet with valid iambic pentameter and correct rhyme scheme. Suddenly the monkeys become extraordinarily productive at generating valid sonnets. Not because they understand poetry. Not because they have developed literary intuition. But because the verification oracle converts their speed and parallelism into outcomes. Fast becomes better when and only when the oracle exists.

This is what happened with the Erdős result. The model did not understand discrete geometry the way a human mathematician understands it. It explored a vast solution space more efficiently than any human or team of humans could, generating candidate constructions and connections at machine speed. The proof checker told it instantly when something was valid. The oracle converted speed into outcome. The monkeys produced a valid proof not because they understood geometry but because someone built a machine that told them immediately when they typed something that worked.

This framing is more honest than the creative insight narrative. It is also not a dismissal. The monkeys produced something real. The proof is valid regardless of the mechanism that generated it. The Erdős conjecture has been disproved and that matters for mathematics independent of whether the model that did it has anything resembling mathematical understanding. Outcomes are outcomes.

But the mechanism matters enormously for what the result tells us about AI capability more broadly. And the mechanism is exhaustive search with binary automated verification, not creative insight.

The Oracle Is the Load-Bearing Structure

The oracle, the automated binary verification mechanism that tells the model instantly when a candidate solution is valid, is the load-bearing element of the entire result. Without it, the speed and parallelism of the model produces nothing useful. With it, speed becomes functionally equivalent to insight from an outcomes perspective.

This is why the result landed in formal mathematics rather than medicine or law or strategic planning. Mathematics has the oracle. A proof is valid or it is not, verified automatically, unambiguously, without human judgment. The oracle is centuries old, extraordinarily reliable, and scales perfectly to machine speed. The model can generate a million candidate proofs per hour and the oracle evaluates every one of them instantly.

Medicine does not have the oracle. A diagnostic hypothesis is not valid or invalid in a way that automated verification can confirm without clinical trials involving real patients over real time. A billion diagnostic attempts without the oracle produces a billion unverified diagnoses. The speed advantage evaporates because verification requires human judgment that cannot be automated at scale.

Law does not have the oracle. A legal argument is not valid or invalid in a way that automated verification can confirm without human judges evaluating actual cases in specific jurisdictions with specific facts. Speed without verification produces plausible-sounding arguments, not verified correct ones.

Strategic planning, organizational management, foreign policy, and the other domains where AGI scenarios require comparable capability do not have the oracle. The verification of a strategic insight requires observing outcomes in the real world over real time, which cannot be compressed by machine speed regardless of how many candidate strategies you generate.

The Erdős result is extraordinary within its domain. It is strictly limited to its domain by the oracle condition.

The Thousand Monkeys Framing Is More Honest Than the Alternative

OpenAI’s creative insight framing serves their narrative interests. A model that makes creative leaps is a model on the path to general intelligence. A model that does exhaustive search with binary verification is a very fast pattern matching system that works in specific constrained domains. The former sells the AGI story. The latter does not.

But the thousand monkeys framing is not just more honest. It is more useful for understanding what AI will and will not be able to do going forward. The question is not whether AI can make creative leaps in some abstract sense. The question is whether the oracle exists in the domain you care about. Where it does, exhaustive search at machine speed will produce genuine outcomes and the creative versus mechanical distinction collapses. Where it does not, speed remains different from insight in all the ways that matter for reliability, accountability, and deployment in high stakes settings.

The thousand monkeys needed a very specific cage. The cage is the oracle. Inside the cage, the monkeys are extraordinarily productive. Outside the cage, they produce noise.

What the Mechanism Tells Us About Creativity

There is a genuine philosophical question buried in here that deserves acknowledgment even if it does not change the practical analysis. If a model generates a valid mathematical proof through exhaustive search, is the proof less valuable than one generated through human insight? Obviously not. The proof is the proof.

But does the mechanism tell us something about whether the model has developed genuine mathematical understanding? This is where the creative insight framing does the most damage. The claim that AI has developed mathematical intuition comparable to human mathematicians is not supported by the exhaustive search mechanism. A model that succeeds by covering more of the solution space faster than humans can is not doing what mathematicians do when they have insights. It is doing something different that happens to produce valid outputs in domains where outputs can be automatically verified.

The distinction matters for the AGI question precisely because AGI requires the insight mechanism to generalize across domains, not just the exhaustive search mechanism to work within domains where the oracle exists. Exhaustive search with binary verification is a powerful tool with specific applicability. It is not a general intelligence. The monkeys in the cage are very productive. They are still monkeys.

The Honest Summary

The Erdős result is real, valid, and significant for discrete geometry. The mechanism that produced it is exhaustive search with binary automated verification, not creative insight in any sense that transfers across domains. The oracle is the load-bearing structure of the result and its presence in mathematics and absence in most other high-stakes domains determines almost everything about what the result tells us about AI capability more broadly.

The thousand monkeys framing is more honest than the creative insight framing and more useful for understanding what comes next. The monkeys will keep producing valid results in domains where the oracle exists. They will produce noise in domains where it does not. The Pattern Matching Conditions identify which domains have the oracle. Everything else follows from there.

Part three of this trilogy asks the practical question: given all of this, does the Erdős result actually change anything about the employment, AGI, and policy arguments? The answer is more nuanced than either the hype or the dismissal suggests.