A commentary published by the WeChat account AIGC从0到1 said a mathematician “believed” he and Claude had found a proof of the Riemann Hypothesis. The public material reportedly includes more than 10 million lines of Lean code, but there was no paper or peer review, the full compilation had not been independently confirmed, and the mathematician said he did not fully understand the mathematics in the latter part of the proof.
Anthropic mathematician Levent Alpöge said he and Claude found a proof of the Riemann Hypothesis, with the method also claimed to extend to the generalized version. He published a four-page explanation and 14 compressed packages of Lean proof files for researchers to inspect, but the claim has not been independently verified. Terence Tao said understanding the Lean-only proof would take time and offered to help communicate with Alpöge, without endorsing the result.
Anthropic mathematician Levent Alpöge announced that he and Claude discovered proofs of the Riemann Hypothesis and its Generalized form, releasing a Lean formalization described as about 17 million lines across 14 compressed archives. Andrew Sutherland and other independent checkers reportedly compiled the files and found they passed a basic logical check, but 36氪 reported that a crucial mathematical check remained outstanding. Hours later, Ravi Vakil announced a competing human-authored GRH draft by Alessandro Masullo, which was also still under review.
Vals AI’s latest Autoform-Hard leaderboard ranks GPT-6.2 Astra first, with 49 of 56 advanced formal-mathematics problems solved and an 87.50% accuracy rate. Claude Fable 5.2 and GPT-6.2 Sol each solved 48 problems and tied for second at 85.71%. According to Vals AI, Astra completed a Lean formal proof of the Smoktunowicz gap theorem without human intervention in 94 minutes, producing 1,794 lines of Lean code and introducing a new “capacity” argument rather than using the original proof’s three theorems.
An analysis argues that successful Lean compilation is only one part of assessing an AI-generated proof. Its four-part framework asks whether the Lean statement matches the paper, whether the code compiles, whether the proof uses unintended axioms, and whether it exploits bugs in Lean’s kernel. Even if all checks pass, the Lean proof may differ from the proof in the paper’s PDF. The analysis reports that OpenAI’s theorem dump contained 722 papers, around 45% with some Lean code, and that about 38 of 290 corresponding entries did not state, let alone prove, the papers’ main theorems. The author also says these checks have not been systematically applied to many papers in the dump and sees no reason to believe they will fail.
Tiger Sniff AI, citing a preprint by Anders Hansen, Fabian Circelli and Alexander Bastounis, reports that part of OpenAI's Navier–Stokes paper differs from the proposition verified by its Lean code. The report also describes Lean 4 kernel vulnerabilities that reportedly allowed false proofs to pass until the Lean team patched them in version 4.33.1.
36氪, citing OpenAI, reported that roughly 10,000 AI agents completed a Navier–Stokes solution in 88 hours, exchanging 2.7 million messages before submitting their answer. A five-member panel later released Lean code and review material for a partial Yang–Mills solution, while noting that the mass gap and a related limit were not proved. The article says OpenAI would not claim the US$1 million prize.
Haocheng Wang and six coauthors submitted the AuXSpec paper to arXiv on November 16, 2026. AuXSpec is a Lean framework that pairs declarative postconditions with executable counterparts, allowing test cases and execution agreement to provide reinforcement-learning feedback for specification generation; the authors also synthesized 3,037 LeetCode problems and expanded training to 19.5K samples through expert iteration on Codeforces. The paper reports that training Qwen3.5-9B and Qwen3.8-27B with this reward raised AuXSpec-Test pass@4 to 82.7% and 84.3%, respectively, and says the authors released the trained weights, synthesized problems, and the expert-reviewed AuXSpec-Test benchmark.
An essay argues that AI has pushed formal verification into the mainstream, citing growing startup activity, a billion lines of Rocq generated for one project, and work by Theorem with Anthropic and Harmonic on software and hardware verification. It cautions that “formally verified” systems can still contain bugs because specifications, models, provers, compilers, and machine semantics may remain trusted or imperfect. The author predicts binary-level proof generation for some first-order properties of single-threaded programs, while concurrent, distributed, higher-order, and non-functional properties will require years more work.
Google expanded its Antigravity platform with Teamwork, a system built from specialist agents powered by Gemini, and introduced the iterative /boost programming workflow. Google says Teamwork formalized work related to the Weil conjectures, reproduced a formal proof about finite-time blowup in the three-dimensional Navier–Stokes equations, and helped resolve the Courtade–Kumar conjecture. The report also says a 30-agent chemistry team reached 95% Top-1 and 100% Top-3 accuracy on a complex-molecule benchmark, while another agent team rewrote Google’s libgav1 AV1 decoder from C++ to Rust.
Google DeepMind and Google Quantum AI reported using Gemini 4-driven AI agents to design quantum programs for three scientific simulation tasks: fusion, photochemistry and drug metabolism. The programs’ correctness and resource costs were checked in Lean, and the team reported lower costs than the previous best solutions, with the largest reduction about 90%. The results were based on resource estimates rather than execution on quantum hardware. The team said it would publish the Lean library, problems and verified programs, but outside researchers could not independently inspect the proofs at the time of the report.
A report says Edward Witten credited Claude Opus 5.5 with proposing a possible mathematical construction and helping with technical details in an arXiv preprint on lattice definitions of gauge theories with chiral fermions. In related preprints, Nathaniel Craig credited GPT-6 Astra with substantial mathematical derivations and Lean-based verification, while Stephen D. H. Hsu said GPT and Claude helped produce technical details and check results. All three papers remain preprints and require further examination by researchers.
Google DeepMind said its AlphaProof Code framework used AI agents and Lean to formally verify specified security properties of compiled Signal libsignal code. Five researchers and agents reportedly produced about 6.5 million lines of Lean proof code in four weeks, and tests with two previously fixed bugs produced machine-checked proofs that the buggy versions violated the requirements. The work covers selected properties of a specific libsignal command-line test binary, not the entire Signal application, group messaging, or every post-quantum property; some underlying cryptographic operations remain subject to assumptions.
David Andrews and Chao Zhang report that AI agents rewrote 4,071 CPython 3.15.2 functions in Lean, producing 613,318 lines, 18,898 theorems and lemmas across 1,539 files. Each rewrite was differential-tested against real Python on tens of thousands of inputs, then checked by a separate agent and Lean’s proof checker. The project also generated 418 bug reports, including about 55 cases said to be capable of crashing the interpreter or corrupting memory.
Import AI reports that Physical Superintelligence (PSI) says an internally developed research system, running a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture. PSI says the system produced a complete proof on the third day of its run and that Lean, a proof assistant that checks each step, accepted it. PSI also says the system improved the result twice while working on related open problems, with Lean accepting the final version. The account comes from PSI and has not received independent expert review in the supplied report.
Anthropic's developer blog says a team of about 1,000 agents running an early version of Claude Fable 5.2 reproduced a formal proof of Fermat's Last Theorem using general-purpose tools, a shared database, and layered safety controls. The Claude Developer Blog describes it as a toy experiment, while a Reddit summary reports that the proof was completed in 30 hours with minimal human intervention. The later proof was not publicly linked, so that timing claim has not been independently verified.
Physical Superintelligence (PSI) says a system powered by an undisclosed Anthropic model produced a proof of the Aaronson–Ambainis conjecture in three days. PSI published the paper, proof code and a Lean verification record, but State of Proof said it had not independently reproduced the check or confirmed peer review.
An analysis argues that the sim-to-real gap may help explain why AI has made rapid progress in programming and mathematics while its real-world impact in many professions remains modest. Programming and mathematical reasoning can be replicated and verified in digital environments, whereas many white-collar professions involve people and physical environments that are difficult to simulate. The analysis uses robotics examples to illustrate the challenge. It also cites GDPval, where the best score reportedly rose from 47.6% for Claude Opus 4.1 when the benchmark was released to more than 85% for current leading models.
Axiom Math announced that its AI agents formalized about 7,654 pages of the Stacks Project into roughly 16 million lines of Lean code, as reported by TechCrunch. The company says the interconnected library includes definitions, lemmas and theorems, with code divided into a human-reviewed core layer and a frontier layer awaiting review. The scale and reuse potential are company claims, and Lean checks logical derivations rather than whether the formalized statements faithfully express the original mathematics.
Anthropic open-sourced FlashProve, a research prototype that uses zero-knowledge proofs to let users check whether a response was produced by a model whose weights were previously committed. The proof does not reveal the weights, intermediate values or which experts a mixture-of-experts model activated. Anthropic says tests using Llama 4 Maverick produced proofs of about 19 MB for 1,024-token requests, while verification took about 17 seconds on a 24-thread CPU without a GPU or access to the weights. The company reported proof generation at about 110 tokens per second on one Nvidia H100 and at 70 to 95 times the GPU time of ordinary inference under its comparison conditions. The prototype supports only mixture-of-experts models and contexts of up to 1,024 tokens; side-channel attacks and formal equivalence between the Lean specification and executable implementation remain unresolved, according to the supplied reports.
Sam Altman said OpenAI is preparing a next-generation GPT that may be the internal model used by about 10,000 model-driven agents in a roughly four-day study of the Navier–Stokes problem. Reuters reporting said the experiment began September 1 and produced a mathematical proof attempt concerning whether a fluid could develop a singularity in finite time under smooth external forces. The report said GPT-6 Astra then used Lean for about 17 hours to formalize and verify the work. The forthcoming model's release date was not specified, and the proof has not received an independent peer-reviewed assessment in the supplied report.
Sam Altman said at ADIPEC 2026 that an upcoming next-generation GPT powered about 10,000 agents in OpenAI’s previously announced Navier–Stokes work, operating collaboratively for roughly four days. According to reports from 36Kr citing Altman and OpenAI’s disclosures, the agents conducted core exploration over about 88 hours before GPT-6 Astra spent another 17 hours completing Lean formalization and verification. These performance details remain company-reported and have not been independently validated in the supplied material.
In an interview with Ben Lorica, Lean creator Leo de Moura discussed how AI is making formal verification more accessible and potentially transforming software engineering. The conversation covered specifications, machine-checked proofs, Rust verification, autoformalization, Mathlib, AI-generated code and autonomous agents, as well as the need to evaluate verification claims and trust proof checkers. They also discussed a vision for a verified software stack.
AI systems have sharply reduced the reported time and cost of formal software verification, while startups including &Truth, Harmonic, Theorem and Axiom Math are expanding the field. &Truth launched publicly with $70 million in funding, and Anthropic and Theorem announced an AI-assisted project that found and patched hundreds of defects in 123 open-source libraries. Stanford researcher Chuyue Sun reported that her AI system completed the remaining specifications and proofs for an encryption library in 11.4 hours at a cost of about $467 after receiving the top-level specification. The reports differ on the human effort involved: WIRED described five months of work by three experts, while MIT Technology Review cited eight months by five major contributors. Axiom Math separately said AI agents produced 43.1 million lines of Lean-checkable formalized code for a mathematical proof in 24 days; that vendor claim was not independently validated. Formal verification remains experimental, and its guarantees depend on the correctness of specifications. Pricing, adoption and its ability to address user errors or hardware failures remain uncertain.
36Kr and QbitAI report that Physical Superintelligence used its in-house Emmy AI research system to produce results on five open problems in quantum information theory. The selected problems come from the Open Quantum Problems catalog and cover entanglement purification, quantum field theory, high-dimensional quantum communication, channel capacity, and quantum causality. PSI says it published papers, Lean formal proofs, and verification code on GitHub for scholarly review, but the reports describe no independent validation.
After nearly seven weeks, mathematician Scott Armstrong said several groups of partial differential equation experts had largely understood the method behind OpenAI’s claimed 166-page Navier–Stokes proof, which reportedly took about 10,000 AI agents and 88 hours. Understanding the approach does not establish that the proof is correct: researchers have identified discrepancies between the paper and its Lean version, while the Clay Institute continues to list the problem as active.
The paper authors present EasyProver, a Lean proving system built with recursive agents, isolated contexts, asynchronous search, parallel tactic probing, and an elastic elaboration service. Using an off-the-shelf frontier model without additional training, they report that it proved 669 of 672 PutnamBench problems, all six IMO 2026 formalizations, 597 of 600 NTP4VC conditions, and all 189 VERINA tasks. The authors also introduce Aegis, which produces 2,045 differentially validated Rust-Lean program pairs and 11,533 candidate specifications for evaluating other provers and supplying reinforcement-learning rewards.
Axiom Math says that dozens of AI agents, coordinated with human mathematicians and engineers, completed a Lean formalization of the first-generation proof of the Classification of Finite Simple Groups in 24 days. The company reports about 43.1 million lines of Lean code and 2.05 million declarations, and has opened the repository for inspection. Axiom also says the formalization uncovered an issue in a published argument for the Unbalanced Group Theorem, leading to an alternative route and the result “Preservation of simple tops.” The repository’s review status is described as a team assessment; external mathematicians still need to verify that the formal statement matches CFSG and independently assess the claimed issue and repair.
A team led by Greater Bay Area University professor Xudisheng reportedly used a Chinese scientific multimodal large model and human-AI collaboration to prove the three-dimensional Ghys conjecture. The process involved seven successive reasoning rounds, expanding a problem graph from 15 to 389 nodes and producing more than 1,000 pages of derivations, according to 新智元. A Peking University team led by Bin Dong and Academician Ruochuan Liu then used the SHEAF framework to formally verify all four main theorems in Lean, involving about 550,000 lines of code over roughly five days with up to 100 agents working in parallel; the paper and formalization code were published on arXiv and GitHub.
A report says GPT-6 Astra advanced the upper bound on prime gaps from the 70 million established by Zhang Yitang in 2013 to 186. The report notes that AI can combine existing theorems, search for parameters and contribute new mathematical proofs. However, the claimed bound of 186 is expressly conditional on the LEAN proof containing no errors.
What problem it solves
AI agents need reliable access to web content and specialized data sources, including pages that require JavaScript rendering. Firecrawl provides search, scraping, crawling, document parsing, page interaction, and data-provider access in formats suited to agent workflows.
Key features
Search and scrape the web — Search the web with full page content, or convert a URL into Ma…
What problem it solves
Kyverno helps platform teams enforce best practices and security standards. It enables security, compliance, automation, and governance through policy-as-code.
Key features
Admission control and background scanning — Kyverno can validate, mutate, generate, and clean up resources using Kubernetes admission controls and background scans, so policies apply both to incoming requ…
What problem it solves
Web teams and developers can need to rebuild an existing site in a modern codebase when migrating platforms, recovering lost source code, or learning how a production site is constructed. This template gives an AI coding agent a repeatable workflow for recreating a site as an editable Next.js application while comparing its appearance and behavior with the source.
Key featur…
What problem it solves
Accumulated caches, build artifacts, large files, duplicates, and application leftovers can make disk usage difficult to understand and manage. Cleaning these areas, removing applications, and adjusting system settings can also risk accidental data loss or unintended changes if users cannot review actions before they occur.
Key features
Deep and targeted cleanup — Locate and…
What problem it solves
Story creation often requires coordinating ideas, settings, characters, continuity, review, revision, covers, interactive state, and multilingual delivery. InkOS addresses this by giving different creative formats a shared workbench and harness that manages context, state, and artifacts.
Key features
Unified production harness — Studio Chat, the TUI, CLI interaction, and pro…
What problem it solves
System maintenance can require removing unnecessary files, managing caches, scanning for malware, and adjusting privacy-related settings across several platforms. Kudu addresses these tasks with open-source desktop tools and a scriptable CLI that run locally.
Key features
Local system cleaning — Clean system, browser, application, and game caches, manage registry and startup…
What problem it solves
Production ontology changes can introduce consequences that are not visible in a text diff, making it difficult to assess their impact before deployment. Reviewers also need evidence they can verify later without trusting the engine or the person who applied the change.
Key features
Plan and control ontology changes — The lifecycle workflow plans changes, reports their blast…
The most direct way to give an Agent long-term memory is to write history into a database and retrieve a few relevant records when a new question arrives. This pipeline soon encounters two different requirements: a personal assistant needs to remember specific people, events, times, and preferences; a task Agent needs to learn from past actions what to do next. The former fears losing details to c…
What problem it solves
3D scans need to be checked against nominal geometry, measured, and understood before they can guide manufacturing or inspection workflows. ScanRuler addresses this with browser-based measurement, deviation analysis, wall-thickness mapping, reverse engineering, and mesh editing while keeping local file processing on the user’s computer.
Key features
Geometric measurement and…
What problem it solves
AI applications need isolated environments where agents can execute code, run commands, manage files, and operate browsers or desktops. Teams also need a consistent way to run these environments locally with Docker or on Kubernetes without switching between separate interfaces.
Key features
Agent working environments — Provide agents with isolated environments for command ex…
What problem it solves
AI coding tools can lose context between tasks and struggle to carry development work forward consistently. Ralph addresses this by starting each iteration with a clean context while retaining progress through git history, progress.txt, and prd.json.
Key features
Fresh-context autonomous iterations — Each iteration launches a new AI instance with clean context. Work continui…
What problem it solves
GitHub Actions workflows need repository files available before they can build, test, or modify code. Checkout also has to handle authentication, fetch depth, sparse working trees, submodules, and trusted workflow contexts without introducing avoidable security risks.
Key features
Configurable repository checkout — Check out the triggering repository or another repository by…
What problem it solves
Voice input and voice output are usually handled by separate cloud services, which means audio and transcripts leave the machine and the two halves of the voice loop do not connect. Voicebox aims to cover both directions locally: cloning voices from short samples, generating speech, dictating into any text field, and letting MCP-aware agents speak back in a cloned voice.
Key…
What problem it solves
AI-generated content can contain provenance marks that remain in text, metadata, or media files even when the content is otherwise ready for use. The project addresses this by inspecting and removing marks across multiple vendor ecosystems and file formats while keeping the agent workflow separate from the processing service.
Key features
Multi-vendor text cleaning — Removes…
What problem it solves
Web scraping workflows often span HTTP requests, JavaScript rendering, browser interaction, data extraction, and large-scale crawling. Scrapling brings these activities into one Python library and supports use from Python, the terminal, or an AI agent.
Key features
Fetch and interact with websites — Fetch static content through HTTP requests or render and interact with dynam…
What problem it solves
VoiceStudio provides voice cloning, voice design, video dubbing, dictation, transcription and audiobook creation in 646 languages. It runs on laptops, desktops, servers, phones, tablets and more.
Key features
Local voice creation — Clone a voice from a clean reference recording, design a voice from a written description, and generate narrated or batch-produced audio.
Source:…
What problem it solves
CLI-based AI agents often cannot reach content that requires authentication, appears inside videos, or lives on platforms with weak general search coverage. They also need clean, structured results from platforms such as V2EX, GitHub, and Xueqiu rather than manually assembled web pages.
Key features
Broad platform coverage — Search and read channels cover Bilibili, V2EX, Git…
What problem it solves
Concat provides a video editor that runs locally, with no watermark, account, subscription, or upload. It supports graphical editing as well as scripting and AI-agent workflows through a JSON-RPC API and MCP server.
Key features
Local editing and AI tools — The editor processes footage locally and includes offline-capable tools for captions, text-to-speech, voice cloning, an…
What problem it solves
Managing multiple MCP servers often leaves client configurations scattered, setup duplicated, and access control inconsistent. Teams can also lack a shared view of available tools, while local setups become difficult to reproduce.
Key features
Unified MCP endpoint — Register MCP servers once and connect multiple clients and custom agents through a single endpoint.
Source: RE…
What problem it solves
Complex robotics and computer-vision processes produce multiple synchronized data streams whose relationships over time are difficult to understand from text logs or conventional debugging tools. Rerun addresses this by recording, visualizing, querying, and streaming multimodal data from sources such as robots, human-data rigs, simulation, and web video.
Key features
Multimo…