719 math manuscripts, two promised open-weight models and another AI safety exit

This week's Last Week in AI maps three different stories about verifiability. OpenAI showed hundreds of mathematical results without releasing the model, two Western teams announced open-weight models without immediately downloadable weights, and a safety researcher described limits in internal risk governance.

OpenAI published 719 manuscripts while keeping the model private

At the time of the roundup, the public repository contained 719 manuscripts grouped into 372 topic families. Some proofs have Lean formalizations, and OpenAI added 10 summaries of the model's reasoning, compute estimates and statistics on attempted problems. The frontier model itself was not released.

OpenAI said it consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study about publication. The group had previously recommended disclosing model names, prompts and compute costs and warned against testing advanced problems on proprietary models that the wider scientific community cannot use.

Mistral and Reflection sell control while the weights come later

Mistral introduced the multimodal Mistral Large 4 with 1 trillion total parameters and 49 billion active parameters. Reflection AI presented the text-only Beam with 501 billion total parameters and 23 billion active. Both companies frame the models as Western alternatives to Chinese open-weight systems.

Neither set of weights was available for download at announcement time. Mistral opened a moderated API and planned an October weights release. Reflection accepted early-access registrations and also promised weights, a technical report and a model card later in the month. Companies and governments may value self-controlled deployment, but for now that remains a promise rather than a verified feature.

A safety researcher's exit puts internal incentives back in view

David Robinson left OpenAI after three and a half years. He said he led drafting of the current Preparedness Framework and oversaw safety reports for 12 frontier launches. In his essay, he criticized a culture that, in his account, improves safeguards only after problems appear.

His argument cannot be checked with a benchmark and represents the view of a departing employee. It nevertheless complements the two technical stories: a repository, model card or open weights alone cannot show how a company resolves the conflict between release speed and caution.

Independent checks must follow each of the three announcements

For the mathematics, the meaningful signals are expert review of the manuscripts and reproduction of formalized proofs. For Mistral and Reflection, the questions are whether the weights arrive, under which licenses and whether independent evals reproduce the reported results.

For safety, the signal is the company's response: whether Robinson's criticism becomes measurable process changes or remains another public resignation. The roundup therefore works as a map of three checkpoints, not as one grand story about the entire market.

Lilith's verdict

There are 719 manuscripts on the table, two sealed boxes labeled open weight beside them and a safety keeper walking out the door. This week delivered no single answer, only three lines for independent inspectors to pursue.