[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics

New developments in AI mathematics show significant advances, with OpenAI releasing 722 mathematical manuscripts from an internal model. These results include claims regarding faster integer multiplication and uniqueness results for problems like the elastic inverse problem. A competitor noted these findings as a major moment in mathematical history, specifically referencing the Quasi-Riemann Hypothesis. The computational aspect involved approximately three hours of ChatGPT Pro thinking compute per result, suggesting that complex solutions can be achieved rapidly by AI systems compared to traditional methods.
Meanwhile, Mistral Large 4 is available with 1 trillion parameters and multimodal capabilities, with open weights promised later. Independent evaluations place it competitively in various benchmarks, scoring well on an Intelligence Index but facing scrutiny regarding cost and the availability of open weights. Other models like Gemma 2 offer multimodal embeddings, and decision models are emerging as a product category. Concerns remain regarding the validation of AI math results, potential verification errors, and the relative performance gaps between different models and external indices.

* OpenAI released 722 mathematical manuscripts from an internal model in a public GitHub repository.
* The release includes papers, proof artifacts, and reasoning summaries derived from evaluating about 4,000 research problems.
* One contributor highlighted a result for integer multiplication faster than n log n.
* Another noted a uniqueness result for the elastic inverse problem, which was open in 3D since 1994.
* Levent Alpöge called results related to quasi-Riemann and no-Siegel-zeros the most significant moment in mathematical history.
* The solutions for Navier-Stokes were achieved in 88 hours using 10,000 agents, corresponding to three hours of average ChatGPT Pro compute.
* Mistral Large 4 has 1T total parameters and 49B active parameters.
* Mistral Large 4 was pre- and post-trained on approximately 3,800 Grace Blackwells.
* Gemma 2 is an open multimodal embedding model built on Gemma 4 with modular encoders.
* OpenAI's Decisions API uses GPT-6 Luna and is up to 10x faster than the Responses API.

The narrative centers on the accelerating speed and scope of mathematical discovery facilitated by large language models, framed as a paradigm shift in reasoning. The most significant pattern emerging is the tension between claimed breakthrough significance—such as solving problems related to the Quasi-Riemann Hypothesis—and methodological skepticism regarding the process of generation itself. The juxtaposition of rapid computational achievement (3 hours for complex solutions) with documented potential errors (20% of results being disproofs or counterexamples) forces a re-evaluation of what constitutes "discovery" in an AI context.
The system exhibits a pattern of self-referential validation, where the results are judged by peers who themselves harbor skepticism regarding the mechanism, such as the discussion around compute framing and generalization gaps. The emergence of new evaluation indices (like the Intelligence Index) and auditing processes suggests an industry effort to impose external structure on internal capabilities, attempting to manage the perception of capability rather than just reporting raw output. The open-weight releases, like Gemma 2, alongside proprietary decision APIs, demonstrate a bifurcated approach: foundational components are being opened for community access while high-stakes application layers remain tightly controlled, creating an emergent tension between open research and commercial deployment.
The critical implication lies in the governance of mathematical truth. If AI can generate complex proofs rapidly, the focus must shift from *what* is proven to *how* we verify the underlying reasoning steps. The pattern suggests that claims of revolutionary impact are leveraged effectively when they touch foundational concepts, demanding a systemic approach to meta-reasoning and accountability before accepting these results as definitive anchors in mathematics.
Bridge Questions: If the process involves disproofs, what new metrics should govern the weight given to AI-generated mathematical outputs? How can we establish trust in reasoning chains that are deliberately obscured by computation time? What are the long-term consequences when expertise in fundamental mathematics becomes partially outsourced to high-compute inference?