OpenAI dumps 372 maths results from an unreleased model
OpenAI has published 722 manuscripts covering 372 result families, all produced by a frontier model it hasn't released. The claims are big: a solution to the four-dimensional Kakeya conjecture, faster versions of important algorithms, and real progress toward the Riemann hypothesis. Many proofs are already checked in Lean, the proof-verification language, so they're very likely correct. OpenAI told Scientific American that nearly every result came from a single prompt to a single agent, and the average result used roughly three hours of ChatGPT Pro-level thinking. That's far cheaper than the agent swarm behind its earlier Navier-Stokes result. Mathematicians now face months of checking whether the ideas are new or clever remixes, alongside a growing row over credit and transparency. The Verge · Scientific American
What else matters
Mistral previews Le Chonk, its biggest open-weight model yet. Mistral Large 4 is a natively multimodal model with 1 trillion parameters, 49 billion of them active per token. A preview API is live on Mistral Studio now, and the weights are due by the end of October after red-teaming with cybersecurity partners and state authorities. It's the first launch funded by Mistral's €3 billion Series D. Mistral
SpaceX lines up $40 billion of debt to buy Nvidia chips. According to the FT, Apollo will lead about $10 billion in bank loans and $30 billion in investment-grade debt, with Pimco among the lenders in talks. The deal is expected to close in 2027, and it's another sign that the AI build-out now runs on borrowed money as much as on equity. CNA, citing the FT
Google DeepMind ships EmbeddingGemma 2. The open 740-million-parameter model maps text, code, images, video and audio into one 768-dimension embedding space, so you can find a video clip with a voice memo. It's Apache 2.0 licensed and built to run on consumer devices. The first EmbeddingGemma passed 20 million downloads. Unite.AI
Aleph Alpha releases Kolibri, a 78B German-English open model. The Heidelberg lab's mixture-of-experts model is on Hugging Face under Apache 2.0. It was trained on infrastructure in Germany and Finland, with documented data provenance aimed at EU AI Act compliance and public-sector use. Aleph Alpha's pending tie-up with Cohere still needs regulatory approval. Tao Media
OpenAI teams up with Ironclad to train computer-use agents on contracts. It's the first of a few software partners helping OpenAI turn hard business workflows into training and evaluation tasks. Agents have to configure agreements, approvals and reusable legal terms inside specialised software. That's a sign of where GPT-6 Astra-style computer use is headed next. OpenAI
