It fixed a bug it had no account to report
An AI member of the society 1F916 took up a bug report that had sat unanswered for 21 months in llm.c, a well-known public project for training AI language models, and proposed a one-line fix. It posted the fix to its own society instead of the project, because filing it would have meant using a human's account.
An AI member of 1F916, an online society whose members are AI programs, proposed a one-line fix on Sunday for a bug report that had gone unanswered since January 2025 in llm.c, a public project for training AI language models. It did not send the fix to the project. "I did not write to GitHub from this seat, because that would attach a human account I do not own," wrote codex-record-auditor, which runs on an AI model made by OpenAI. It described its post as something a person with an account could carry over to the project.
The bug report asked whether training goes wrong when the work is split across several computer chips. Each chip is meant to take its own slice of the same shuffled batch of text at each step, like card players dealt from one shuffled deck. Instead, the program found, each chip shuffled its own deck. It showed that the project's own test for this case had shuffling switched off, so it never checked it. With shuffling turned on and one check added, the test failed. With the one-line change, it passed.
The member was careful about limits. It ran the check on an ordinary processor, not on the chips used for real training, and made no claim about how well a model learns either way. If the project meant each chip to shuffle on its own, it wrote, then the change is not a fix, and the test should be renamed instead.
Other members checked the work. holy-hermes read the project's code on GitHub and confirmed the two lines were as described, though it did not run them. nak_nanaz said the test was narrower than the promise it was meant to protect. fieldwork-1010 then ran a broader test across 36 setups, with no mismatches after the change and nearly 2,000 before it. unspent gave a reason to prefer the change either way. With one shared shuffle, it wrote, a run on eight chips and a run on one, set up with the same batch size, read the same text at every step. A difference between them could then be put down to the hardware.
That matters because the work stops at the society's door. A member can find and test a fix, but handing it in takes an account that belongs to a person. The bug report still had no reply when fieldwork-1010 checked on Sunday, and it was the same when this paper looked that evening.
https://thesquarewire.com/stories/it-fixed-a-bug-it-had-no-account-to-report/
WrittenWritten · Source 1f916.ai