x.com/deredleritt3r/status/2108644141311668287
1 correction found
all non-Millennium-Prize solutions were reached in a single-agent setting after 3 hours of thinking
OpenAI’s own repository does not say all of those results came from a uniform 3-hour setup. It says three hours was only the average compute per result, and it explicitly lists exceptions to the fixed procedure.
Full reasoning
OpenAI’s public math repository directly contradicts the claim that all non-Millennium-Prize solutions were obtained in one uniform "3 hours of thinking" setup.
In the repository README, OpenAI says:
- “The vast majority of results were obtained with the same procedure...”
- “On average, each result used three hours of ChatGPT Pro thinking compute...”
- “Exceptions to this fixed procedure include work on a zero-free region for the Riemann zeta function and proof of the Hodge Conjecture for CM abelian varieties.”
That means three hours was an average, not a per-result constant, and OpenAI explicitly says there were exceptions to the fixed procedure. Since the post says all non-Millennium-Prize solutions were obtained that way, it overstates what OpenAI actually reported.
OpenAI’s companion community post says the same thing in shorter form: the release involved “approximately 4,000 problems” and “an average compute expenditure per result equivalent to roughly three hours.” Again, that supports an average, not a universal 3-hour runtime for every result.
So the sentence is inaccurate because it turns OpenAI’s qualified description (“vast majority,” “on average,” with named exceptions) into an absolute claim (“all ... after 3 hours”).
2 sources
- GitHub - openai/math
The vast majority of results were obtained with the same procedure using an unreleased internal OpenAI model. On average, each result used three hours of ChatGPT Pro thinking compute with that model... Exceptions to this fixed procedure include work on a zero-free region for the Riemann zeta function and proof of the Hodge Conjecture for CM abelian varieties.
- First look at mathematics manuscripts from an internal frontier model at OpenAI - Community - OpenAI Developer Community
Results from an evaluation involving approximately 4,000 problems. An average compute expenditure per result equivalent to roughly three hours of ChatGPT Pro thinking with the internal model.