Last Week in AI #346 - 719 math manuscripts, 2 Western open models, 1 more safety resignation

OpenAI has published 719 mathematical manuscripts containing proofs produced by one of its unreleased frontier models, according to the newsletter's roundup. The release puts a large body of machine-generated mathematics into public view from a model that is not itself available.
The same roundup notes two new open-weight model launches, from Mistral and from Reflection AI, framed as an effort to rival Chinese open models. It also records another resignation tied to AI safety.
The proof output is the most concrete item, and it is worth reading as an evaluation artifact rather than a model release: 719 manuscripts is a volume that suggests the model was run across a broad set of problems rather than a curated handful. Proofs are unusually checkable work, since a mathematical argument can in principle be verified line by line, so publishing them offers outside mathematicians something to inspect directly. This suggests the value lies in whether readers find correct, novel or merely plausible-looking arguments, and verification is the hard part.
On the open-weight side, the source names the two companies and the competitive framing but gives no parameter counts, licences, benchmark results or availability dates. In practice, open-weight releases matter to developers because the weights can be downloaded and run or fine-tuned in-house, unlike API-only models.
Why it matters: developers evaluating model capability now have a large public sample of frontier-model math output to study, and teams weighing self-hosted options have two more open-weight candidates to test, though both need independent benchmarking first.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
OpenAI publishes hundreds of math proofs from unreleased frontier model, Mistral and Reflection AI launch open-weight models to rival China, and more!