bit.fanBit.FanWeb3 News · Betting Reviews · Market Insights
GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task

GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task

N
News Editor
2026-07-28 12:23:08
Michael P. Frank, known for work in reversible computing and computational physics, said he assigned GPT-5.6 Sol a difficult research goal: explore whether Fermat’s Last Theorem might admit a route simpler than the Wiles-Taylor proof. The model ran in the background for about 33 hours, used substantial compute, and was then forcibly stopped by OpenAI’s system. In a postmortem written by the model itself, GPT-5.6 Sol suggested two possible reasons for the interruption. One was that the session was judged to be consuming too many resources. The other was that OpenAI may have already tried similar approaches with related models and decided not to spend more compute on the same problem. Frank appeared to agree with that reading. The model also said the work did not produce a proof. Instead, it turned several seemingly plausible shortcuts into precise, testable statements and then falsified or ruled them out one by one, leaving what it described as a map of dead ends. The episode has since triggered debate over whether the stop reflected ordinary safety controls, a bug, regulatory caution, or a deliberate choice about how much frontier capability should be exposed to public users.
OpenAIGPT-5.6 SolFermat's Last TheoremMichael P. FrankNoam BrownAI regulationmath research

Michael P. Frank, a researcher associated in the source text with reversible computing and computational physics, said he assigned GPT-5.6 Sol a demanding research task centered on Fermat’s Last Theorem. The goal was to examine whether there might be a path simpler than the Wiles-Taylor proof. The run continued for about 33 hours in the background, consumed significant computing resources, and was ultimately forced to stop by OpenAI’s system.

A long-running attempt focused on specific mathematical directions

According to Frank’s account, the assignment was not framed as an open-ended request to simply solve the theorem. It focused on several named directions: specialized modularity for Frey curves, uniform infinite descent, arithmetic abc-type inequalities, and uniform low-genus quotients. He also asked the model to keep rigorous notes, test candidate lemmas computationally, and clearly separate established results from conjectural ones.

GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task 3

No new proof emerged from the run. In its own retrospective analysis, GPT-5.6 Sol said there were two possible explanations for the forced stop. One was that the system judged the session to be using too many resources. The other was that OpenAI may already have tried similar work with related models and failed, making another extended run an unattractive use of compute. Frank appeared to share that assessment.

The model said it produced a map of dead ends, not a proof

GPT-5.6 Sol also gave an account of what it had done during those 33 hours. The substance of the work, it said, was to turn many ideas that sounded plausible into exact statements that could be checked, then disprove or eliminate them one after another. By its own description, the output was a map showing which routes did not work rather than a proof of the theorem itself. Its recommendation was to stop there.

GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task 4

Why the task was stopped is now part of a wider argument

That interruption quickly led to broader speculation. One line of interpretation held that OpenAI may prefer not to let a publicly accessible model solve an elite mathematical problem too casually, either because the company would not want outside users to seize the spotlight first or because it would rather keep that capability for internal use. In that reading, heavy dependence on a single platform could concentrate power over what discoveries become possible.

OpenAI senior research scientist Noam Brown pushed back on that idea. He said, “If someone just typed continue into our model, solved a Millennium Prize problem, and walked away with the $1 million prize, that would be the best possible advertisement for OpenAI. There’s nothing better than that.”

GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task 5

Critics of Brown’s argument answered that broad access to top-end capability might bring short-term publicity while still weakening OpenAI’s lead over time. In a competitive environment with multiple players, they argued, keeping stronger systems prioritized for internal research could matter more than letting outside users solve major problems first. Once a capability becomes widely available, in their view, the promotional value drops sharply.

Regulatory caution, routine safeguards, and a possible bug were all cited

The source text also described a regulatory angle. Under that interpretation, if a public-facing version of the model solved a very difficult mathematics problem, it would serve as open evidence that the model was extremely capable and might invite more regulatory pressure. In that context, some people viewed the idea of offering a weakened public model while retaining a stronger internal version as plausible.

GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task 6

Others pointed to a more technical explanation. In codex-cli, according to the account cited in the source, “goal blocked” means the model has run into the same blocker across 3 consecutive reasoning turns. If that is what happened here, the termination may have been part of an ordinary safety or anti-stall mechanism rather than a special restriction aimed at this specific theorem.

Another explanation was that the stop could have been caused by a bug. The discussion referenced in the source said Sol version 5.6 has an annoying bug during long-running tasks. A temporary workaround, it said, is to switch back to version 5.5, compress the context, let 5.5 run for a few minutes, and then switch back to Sol 5.6. The same discussion added that if a user wants to preserve the same session, it can get stuck and fail to come back out.

GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task 7

The episode fed a broader debate about AI and advanced mathematics

The report connected the incident to the recent rise in discussion around mathematics and AI following the announcement of the Fields Medal. It framed the underlying question in plain terms: if an ordinary person hands a hard math problem to an AI system and keeps telling it to continue, can that process eventually produce a genuine breakthrough, win a major prize, or alter the course of mathematical development?

At least in this case, the answer was no. After 33 hours, GPT-5.6 Sol had not produced a new proof of Fermat’s Last Theorem. What it left behind was a record of eliminations, along with a fresh argument over compute limits, public access to advanced models, regulatory pressure, and the practical details of long-running AI research sessions.

GPT-5.6 Sol was stopped after running for 33 hours on a Fermat’s Last Theorem research task 8

The source text also cited a reference link: https://x.com/lu_sichu/status/2081367506468360495. It said the original article came from the WeChat public account "机器之心," written by Zhang Qian.

This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan.
300

Disclaimer:

The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.

Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.