AI model advances in mathematical research
OpenAI said its unreleased artificial intelligence model, Astra, generated solutions to 10 mathematics problems that had remained unresolved for at least a decade across cryptography, quantum computing, geometry and graph theory. The company reported that human researchers prepared the results into research papers with assistance from the same model, with each proof verified through the Lean proof verification system and accompanied by reasoning narratives generated by Astra.
According to OpenAI, the model produced new results on high-dimensional sphere packing, binary and spherical codes, arithmetic circuit complexity, the closest vector problem, quantum parallel repetition and Ehrhart's volume conjecture. The company also said Astra established the existence of non-sofic groups, produced a disproof of Connes's rigidity conjecture, and resolved three open problems posed by mathematician Paul Erdős in extremal graph theory and Ramsey theory. OpenAI estimated the solutions would have cost about $2,000 to generate using its Sol API pricing.
Vibe coding expands software creation
The practice of "vibe coding," building software by describing intent to an AI rather than writing each line manually, is moving beyond professional developers. In a New Atlas piece from August 1, a contributor described using AI-assisted workflows within VSCodium, the open-source version of Microsoft's code editor, to build a custom wake-word system for a Home Assistant voice assistant named "Yaffle," a project that would traditionally have required machine learning expertise involving TensorFlow and speech datasets.
The piece frames a longer arc in which machine code gave way to programming languages, then frameworks and drag-and-drop builders, and now intent-driven development, with the distance between inspiration and working software shrinking. Experienced developers quoted in the trend's broader discourse said vibe coding can produce code its creators do not fully understand, prompting recurring calls for AI-generated code to be disclosed and for closer attention to security and maintainability.
New programming models target multilingual workflows
MiniMax announced M2.1, an update to its M2 model that the company said focuses on improved performance across real-world complex tasks, particularly in multilingual programming and office scenarios. The release emphasizes stronger capabilities in Rust, Java, Golang, C++, Kotlin, Objective-C, TypeScript and JavaScript, with MiniMax describing the model as reaching industry-leading performance across that span.
According to MiniMax, M2.1 strengthens native Android and iOS development, improves the model's design comprehension and aesthetics for web and app work, and introduces upgraded "composite instruction constraints" for office tasks. The company positioned M2.1 as interoperable with several agent frameworks, including Claude Code, Droid, Cline, Kilo Code, Roo Code and BlackBox, and reported benchmark gains on SWE-bench Verified alongside improved multilingual standing relative to Claude Sonnet 4.5.
Researcher caution and verification
Some researchers cautioned that strong math performance does not translate into broader generative AI reliability. Cognitive scientist and AI researcher Gary Marcus wrote on X that while Astra was "apparently" strong at mathematics, that did not mean it would avoid hallucinations, read PDFs properly or follow hard rules. OpenAI has not publicly released Astra, and the results are expected to be assessed by the wider mathematics community through peer review and independent verification, following on from a May disclosure about a separate unreleased model that produced a claimed disproof of the Erdős unit-distance conjecture during testing.
Share this article
