OpenAI floods GitHub with AI math proofs

OpenAI turned math into an info-dump, publishing hundreds of results from an unreleased internal model straight to GitHub and daring the field to keep up. Underneath the headline, Europe got back in the ring with Mistral's 1T-parameter "Le Chonk," Google shipped a multimodal embedding model, and "decision models" quietly hardened into a product category.

OpenAI dumps hundreds of AI-generated math proofs on GitHub

OpenAI published a large batch of mathematical results from an internal frontier model straight to a GitHub repo rather than journals, with Lean formalizations for machine-checking. Outlets count roughly 372 families (Latent Space cites 722 manuscripts from ~4,000 attempted problems); OpenAI says the average result took about three hours of ChatGPT Pro thinking compute, mostly from a single prompt to a single agent. Claimed highlights include a quasi-Riemann Hypothesis result and progress on Birch-Swinnerton-Dyer and Barnette's Conjecture, but OpenAI stresses the model stays unreleased and the results are not independently verified.

Why it matters: This is a deliberate bet that AI output volume now exceeds the math community's capacity to review it — and that Lean verification, not peer review, is the throttle. Mathematician reaction ranges from 'most significant moment in mathematical history' to a 25-Fields-medalist warning letter, so treat the breakthrough framing as contested.

Mistral Large 4 'Le Chonk' lands: 1T total, 49B active, weights promised end of month

Mistral released a preview of Mistral Large 4, a natively multimodal MoE with 1 trillion total and 49 billion active parameters, trained on ~3,800 Grace Blackwell chips in Europe. It is API-only for now at $1.36/$4.18 per million input/output tokens; Mistral says open weights ship at the end of October after safety testing. Independent scoring from Artificial Analysis puts it at 38 on its Intelligence Index — roughly six months behind the frontier per Simon Willison, and still trailing GLM-5.3 on that index — though Mistral claims cyber and vision strengths and a #2 finish in a blind Surge coding review.

Why it matters: A credible non-Chinese open-weight contender matters for anyone who wants auditable weights without a China-origin model, but the preview is API-gated and the headline cyber score partly reflects fewer refusals than rivals. Judge it when the weights actually drop.

Google's EmbeddingGemma 2 unifies text, code, image, video and audio in one 740M model

Google DeepMind released EmbeddingGemma 2 under Apache 2.0, a 740M-parameter natively multimodal embedding model built on Gemma 4 that maps text, code, images, video and audio into a shared 768-dim space. It is modular — 270M for text-only, with loadable 170M vision and 300M audio encoders — supports Matryoshka truncation down to 128 dims for up to 6x storage savings, has an 8K context, and runs in ~191MB RAM for text on a phone. Google reports a ~9.9-point MTEB Code gain over its predecessor, with day-zero support across llama.cpp, vLLM, Ollama, Unsloth and WebGPU.

Why it matters: Embeddings are the one place a closed, hosted-only model is genuinely risky — re-embedding millions of stored vectors when a vendor sunsets a model is expensive. An Apache-2.0 multimodal embedder that runs on-device makes offline RAG pipelines practical and portable.

Decision models harden into a product category

Weeks after TypeSafe AI's Jev, probability-out 'decision models' are proliferating. OpenAI's Decisions API hit public beta on GPT-6 Luna, returning predicates, choices or scores at $0.10/M input tokens with no output charge; Simon Willison shipped an llm-openai-decisions plugin against it. Separately, Musubi released PolicyLM-1.7B, an open-weight decision model for real-time content moderation that applies a plain-English policy in under 50ms without retraining when rules change.

Why it matters: These models trade free-text flexibility for speed and cost by constraining output to a fixed choice set — useful for routing, moderation and policing agent behavior. Watch the cost: one tester found similar accuracy to a full LLM at a fraction of the price, but others argue routing-by-decision-model is overkill for picking intelligence levels.

OpenTPU: an AI-designed open-source accelerator that runs its own inference

A Show HN project, openTPU, puts a full AI accelerator in one readable monorepo — SystemVerilog RTL, a custom ISA, a bit-exact simulator, a kernel language and compiler, and host software driving a real PCIe FPGA card (Xilinx Kintex-7). The design runs ten modern models with real weights on a ~$ scavenged Inspur card, producing tokens bit-for-bit identical to the simulator, and streams MoE experts from host storage for models larger than the card's 4GB. Decode is DRAM-bound at 82-85% of DDR3 peak; the authors frame it as both a research artifact and a teaching tool.

Why it matters: It's a rare end-to-end, auditable look at how an accelerator actually works, from a Python matmul down to the wires — and a concrete data point on how far AI agents can get at hardware design. Good reading for anyone curious about inference bottlenecks beyond the GPU.

Nathan Lambert: the open-model cyber-risk debate is a lose-lose

In a long Interconnects essay, Nathan Lambert argues the discourse around open-weight cyber risk is broken: banning open models while leaving frontier closed-model APIs public would widen the offense-defense gap, since documented attacks to date have mostly come from closed models. He notes that over a month after GLM-5.3's weights shipped — the model Anthropic flagged as a threshold cyber threat — there's little public evidence of the predicted step-change in harm, making the fear-mongering a falsifiable and so-far-unsupported prediction.

Why it matters: This is the counter-case to the vendor reports driving potential open-weight bans, and it's framed as a testable claim rather than vibes. If you build on open weights, the policy fight over their legality is downstream of exactly this argument.

Atlassian wires GPT-6 into Jira, Confluence and Rovo

Atlassian and OpenAI expanded their partnership to power agents across Atlassian's platform and its Rovo assistant with GPT-6-family models, drawing on Atlassian's 'Teamwork Graph' context layer linking projects, docs and decisions. Atlassian says more than 3,000 of its own developers use Codex across terminals, IDEs and code review, and the companies are exploring deeper Jira integrations to assign work to AI agents and track results. OpenAI, in turn, continues to run its internal workflows on Jira.

Why it matters: Enterprise context graphs are becoming the moat for agent usefulness — the model is commoditized, the wiring into your tickets and docs is not. If your org lives in Atlassian, this is the plumbing that decides whether agents see real project state.

Also worth a look

Browse previous days →