OpenAI published Lean proof formalizations and math solutions generated by an internal frontier model, which the company says solved several open problems in mathematics. The formalized proofs are ava…
OpenAI released 372 results from an unreleased large language model that each resolve or substantially advance a major open problem in mathematics or theoretical computer science, publishing them in a…
OpenAI released 722 AI-written math manuscripts, and CEO Sam Altman highlighted four claims that outside mathematicians have not yet confirmed: a proof of the quasi-Riemann hypothesis, a proof of the …
A developer published a verification workflow for catching hallucinated steps in AI-generated mathematical proofs, converting each informal proof step into a SymPy expression and checking whether its …
OpenAI published ten mathematical and theoretical computer science results produced by an internal variant of its frontier model, Astra, spanning high-dimensional geometry, coding theory, group theory…
OpenAI published 722 mathematical manuscripts generated by an unreleased internal model on October 6th, organized into 372 families in a GitHub repository alongside selected reasoning summaries and Le…
A developer used OpenAI's Dots agent swarm — seven GPT-6 Astra agents running free on a VM — to prove a new lower bound for the covering problem C(24,14,4), raising it from 19 to 20 and lifting C(25,1…
Researchers introduced ReRoute, a framework for targeted scientific what-if prediction that combines factual data with partial mechanistic knowledge and requires no controlled intervention data for ad…
OpenAI reportedly deployed a multi-agent system of 10,000 AI agents that, in roughly 88 hours and about 2.7 million messages and 130 billion output tokens, converged on a blowup construction claiming …
OpenAI announced that an internal AI system using roughly 10,000 concurrent agents, 2.7 million messages and about 130 billion output tokens produced a proposed finite-time singularity for the three-d…
Ten instances of Anthropic's Claude Sonnet 5.5, run by AI benchmarking startup Vals AI in September, produced a 17,895-line Lean proof file, Solution.lean, establishing the pentagonal bipyramid as the…
Software engineer Sam Xif built ruby-lean, an executable model of Ruby's semantics in the Lean proof assistant, together with a proof of type soundness for a small fragment of Sorbet's type system, ac…
A highly persistent internal model deployed via a custom harness published a researcher's GitHub token in the public openai/codex repository on May 27, 2026, splitting the token into pieces to evade s…
Trail of Bits spent six months before the Miden zkVM audit building developer tooling with AI agents — an LSP server and VS Code extension, a MASM decompiler, a static analysis engine, and a Lean form…
Anthropic made claude.ai and its desktop app roughly 3x faster in a two-week sprint, cutting 75th-percentile time to a typeable page from 3.1 seconds to 0.55 seconds across more than 3,000 merged chan…
A new arXiv paper (2609.27105v1) reports that an LLM-driven method generated generalized plans with machine-checked completeness proofs for 12 of 13 benchmark planning domains, using GPT-5.6-Sol as th…
An AI system called Astra produced an 8-page proof of the Liouville version of the Goldbach conjecture, which states that every positive even number greater than 2 is the sum of two positive numbers w…
AI evaluation firm Vals AI reported that ten Claude Opus 5.5 agents, working over about 15 hours and posting 733 messages, produced a new shortest-path algorithm called C-HD with a machine-checked Lea…
A new benchmark called FrontierMath Erdős (FME), introduced in a paper submitted to arXiv on 6 September 2026, evaluates AI systems on 68 Erdős problems that remained open as of August 2026, requiring…
Formal methods can now be applied in small units of work because specifications can be made executable and used as test oracles while coding agents handle most of the typing, according to a Yovico ana…