Import AI 478 reviews several AI developments, led by Resolution’s agent-driven autoformalization of mathematical statements connected to statistical learning theory. Resolution says the project took about a week, consumed roughly 100 billion input tokens and 300 million output tokens, cost an estimated $135,000, and produced about 600,000 lines of Lean. The issue also reports Claude Opus 5.5 reaching 18.5% on ProgramBench, Mirendil-Base being trained by one researcher in 3.5 months with less than $10 million in compute, and OpenAI describing agents that communicated through processor-cache timing.
OpenAI announced three claimed results produced by its internal models: a proof of the local smoothing conjecture in harmonic analysis, a refutation of the Exponential Time Hypothesis (ETH) in complexity theory, and a physics result concerning quantum chromodynamics and the Standard Model. The company described the manuscripts as raw research data requiring outside scrutiny. Following advice from external experts acting as “shepherds,” OpenAI said it would release the math output openly, make the computer-science output public while forming an external collaborator group, and leave the physics release to outside experts. The math and computer-science proofs were also translated into Lean, according to OpenAI. Researchers contacted by The Verge welcomed the smaller set of results but questioned restricting initial access to OpenAI-selected experts. Brendan Hassett said the math conclusion appeared well supported but that he had not verified its Lean formalization.
A Deep Tech-edited report says OpenAI released 722 manuscripts generated by AI models, covering 372 sets of problems in areas including number theory, algebra and geometry. OpenAI claimed that some results addressed longstanding open problems and that generating one result required average compute equivalent to about three hours of ChatGPT Pro thinking. The report does not provide independent verification of those claims. University of Texas at Austin computer scientist Scott Aaronson called the event a “Mathocalypse,” while MIT mathematicians Pavel Etingof and Slava Gerovitch argued that AI could free mathematicians from publication competition.
According to Fudan Youth, OpenAI publicly released 722 mathematics research manuscripts generated by an internal model on October 6, 2026, grouped into 372 sets spanning multiple mathematical fields. At launch, only one manuscript had passed manual verification and 42% had passed formal verification, according to mathematician Li Zhiyuan. OpenAI withdrew three Hodge-conjecture-related manuscripts several days later; Li said their errors went beyond the sign error OpenAI had claimed. He also said some cross-domain proof strategies, including transferring tools between fields, were not rigorous and noted that scholars of Lean argued the same conclusion should be verified in multiple formal ways to ensure accuracy.
InfoQ 中文 AI&大模型 reported on the first ICAAS 2026 roundtable, where Michael Levitt, Jack Dongarra, mathematician Xiao Ma and Jiy an He discussed AI’s technical paths, research methods and potential next phase. Participants noted that foundation models are converging on large Transformer architectures, while arguing that research should also explore alternative routes, autonomous learning environments and verification. They disagreed over whether another AI winter could emerge, with concerns including overpromising, token costs, latency and the reliability of autonomous problem-solving.
MIT mathematician Dor Minzer and graduate students Yumou Fei and Shuo Wang posted a 95-page preprint on September 14 after spending three days preparing their seven-year research on the 4-to-1 Games Conjecture. OpenAI released 722 mathematical manuscripts on October 6, including stronger Unique Games and 2-to-1 Games results, but the claims had not received full academic verification, and three manuscripts were withdrawn after a sign error. Minzer's team updated its paper on October 8.
Microsoft Research reports that two researchers, including Xinzhi Zhang of its Machine Learning and Optimization group, have proved the Kannan–Lovász–Simonovits (KLS) conjecture, a question in high-dimensional geometry and probability proposed in 1995 that asks whether the best flat or hyperplane cut is nearly as effective as the best cut of any shape, up to a universal constant regardless of dimension. The blog says the proof establishes a dimension-free bound, improving on a July 2026 result whose bound still grew slowly with dimension, and that the final step repeated a recursive refinement so accumulated errors stay bounded by a universal constant. It describes the work as emerging from a human-directed agentic process in which AI systems explored more than 100 candidate approaches while researchers pruned and redirected the search, and says a separate group of experts posted a full parallel proof the same day, with machine-checked formal verification in Lean planned after final review.
According to a 36Kr repost of a New Zhiyuan article, OpenAI claimed on October 6 that its AI proved a quasi-Riemann hypothesis, excluding Landau–Siegel zeros, drawing positive comments from number theorist Hector Pasten and others. The report says Cureton and Thorner posted a Hexagon preprint on October 12 applying the tools to finite-order Hecke L-functions to obtain a zero-free region to the right of 1-1/6000, which an open-source community then pushed to 1/32, with a Lean formalization generated by the math agent AxiomProver. It adds that a preprint by Frei, Somos and Widmer on arXiv used these results to prove unconditionally the 2007 Ellenberg–Venkatesh bound on ℓ-torsion in class groups, and that the paper states no artificial intelligence tools were used. cnBeta's republished headline says AI again broke the generalized Riemann hypothesis, but the described work concerns zero-free regions rather than the full GRH.
The Simons Foundation announced on October 15 a partnership with Anthropic and OpenAI committing $15 million in combined funding and resources to help the mathematics community navigate the rapid rise of AI, with the foundation managing and awarding the money, including through an open call for proposals, for workshops and initiatives. OpenAI said it is contributing $5 million and will give Simons-supported researchers access to its models and computing resources, while Anthropic said it committed $5 million of funding, compute and Claude access; a 智东西 report said Anthropic's specific contribution amount had not been disclosed. Foundation president David Spergel said the community, not any single company or institution, should drive discussions of how these tools are evaluated, adopted and used.
On September 8, 2026, OpenAI said its internal AI system produced a 166-page Navier-Stokes proof in about 88 hours using roughly 10,000 agents, published a Lean formalization, and attributed the approach to Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa. Several expert groups told Scientific American that the details instead show a different method, with EPFL mathematician Gonzalo Cao-Labora saying the similarities are only surface-level. OpenAI researcher Noam Brown estimates the multi-agent setup contributed under 10% of the success; NYU mathematician Tristan Buckmaster has accused OpenAI of building on work he did with Anthropic's Levent Alpöge, which OpenAI denies, and the Clay Mathematics Institute has been reviewing the proof.
Shanghai AI Laboratory announced that a team led by Xu Disheng, a visiting scholar at the lab and professor at Greater Bay University, used its Intern-S2 science multimodal model in a human-machine collaboration to complete a proof of the three-dimensional Ghys conjecture, an open problem in complex dynamics posed by Étienne Ghys in 1995. A Peking University team led by Dong Bin and Liu Ruochuan used its own SHEAF framework to formally verify all four main theorems of the paper, with the proof passing Lean checks. The paper is on arXiv and the formalization code is public on GitHub; the lab calls it the first time Chinese scientists have used a Chinese model to crack a world-class math problem.
In an October 9 talk at Caltech's Beckman Auditorium, Fields medalist Terence Tao retitled his lecture from "Machine-Assisted Proofs" to "Math 2.0," arguing that frontier AI has moved mathematics from an era of proof scarcity into one of proof abundance. The report relays claims that OpenAI coordinated roughly 10,000 AI agents to produce a Lean-verified finite-time singularity proof for the Navier-Stokes equations in 88 hours and published 700-plus AI-generated math manuscripts covering 372 results on GitHub, that Anthropic said Claude raised the lower bound for nontrivial zeros of the Riemann zeta function from 41.6% to 67.2%, and that Google DeepMind's AlphaProof has reached IMO gold-medal level. Tao warned via a thought experiment about an AI-designed "perfect cancer drug" that correct but incomprehensible output is a dangerous misalignment, and proposed Math 2.0: human-AI collaboration that feeds back into human understanding, such as tackling "open explanation problems," merging papers into an evolving knowledge network and building large formalized math knowledge bases.
Google DeepMind’s AlphaProof Nexus, a Gemini 3.1 Pro-based system that drafts Lean proofs and uses compiler feedback to revise them, solved 9 of 353 open Erdős problems, including two that had stood for 56 years, according to a study published in Science. The system also proved 44 open OEIS conjectures, settled a 15-year-old algebraic geometry question and improved a known bound in min-max optimization, with inference costs of a few hundred dollars per problem. DeepMind says a simpler LLM-plus-Lean loop also solved all nine Erdős problems; an independent audit cited by 36Kr found one solution was already implied by an earlier theorem, which would reduce the count to eight under a stricter definition.
A team led by Princeton scholar Ayush Khaitan, working with NVIDIA's Humanfia team, says it has completed a full Lean formalization of the Thurston geometrization conjecture. The project spans roughly 4.7 million lines of code, about 2.7 million of which were added in around two weeks on top of roughly 2 million lines of prior work by Chow, Liao and Qin, and it has been open-sourced on GitHub. The team includes Bennett Chow, a Yau Shing-Tung student and UC San Diego professor, plus formalization leads Ziyang Qin and Yuan Liao; Stanford mathematician Jared Duker Lichtman publicly called the work amazing. The report notes that the geometrization conjecture is broader than the Poincaré conjecture, which Perelman proved as a byproduct, but NVIDIA has not disclosed the details of AI's role, and much of the article's account is inference.
According to a 36Kr repost of a New Zhiyuan article, OpenAI published 722 math papers at once, said to include a counterexample showing failure of solutions to the Navier-Stokes equations and other claimed results on well-known open problems; the report also notes the papers still await verification. On Bloomberg's Odd Lots podcast, MIT associate dean of engineering education Justin Solomon said the proof is dense with technical detail, counterintuitive inequalities and extreme bounds, and that he stalled after two pages, able only to say it is "probably right." Fields Medalist Terence Tao, in a Caltech "Math 2.0" talk, offered a thought experiment about an AI-designed cancer therapy that passes formal verification and trials but whose mechanism no human understands, raising questions of interpretability and alignment.
Fields Medalist Terence Tao argues in a new talk and essay that AI is turning mathematical proofs from scarce to abundant, ending the "Math 1.0" era built on proof scarcity and opening what he calls Math 2.0. He attributes math's early exposure to AI to three overlapping factors: proofs can be objectively checked, mathematical objects and reasoning are highly digitizable, and much high-quality literature already exists in digital form — yet most problems at the current frontier remain unsolved, and vendors tend to show only cherry-picked successes. Tao calls for more transparent evaluation that reports failures and compute costs alongside results, for less emphasis on being first to produce a proof and more on explanation, review and knowledge organization, and says the SAIR Foundation he co-founded will soon release open math models aimed at everyday research needs.
Liam Fedus, a ChatGPT co-creator and former OpenAI post-training lead, told the podcast The Frontier, in conversation with Mercor founder Brendan Foody, why his startup Periodic Labs is building its own physical laboratories: math has judges like Lean that a compiler can check line by line, but questions in materials science, chemistry and physics can only be settled by real experiments, while web data is noisy, papers omit key details and negative results are almost never published. He said Periodic combines language-model reasoning, computational physics simulation and high-throughput lab work, starting with materials discovery, and that the company's first model, the roughly 1-trillion-parameter Periodic Neon released in September, beats GPT-6 Astra on its own benchmark — a company-reported result. He also referenced OpenAI's release of 722 math manuscripts from an unreleased internal frontier model, and discussed ChatGPT's shift from a completion interface to a chat interface and RLHF reward models' insensitivity to correctness.
In a guest post on Terence Tao's blog, Thomas Hales surveys formalization and the Lean theorem prover, arguing that autoformalization became a practical reality in 2026: milestones cited include Math Inc.'s autoformalization of the 24-dimensional sphere-packing problem, Meta's ATLAS textbook project, Anthropic's 13 million lines of Lean for Fermat's Last Theorem in 11 days, and OpenAI's Lean formalization alongside its Navier-Stokes announcement. He also examines Lean's 2026 "Summer of Soundness Bugs" — one bug yielded an illicit disproof of the Collatz conjecture — which were found by security researchers using frontier models, quickly patched, and followed by re-verification of mathlib. Hales stresses that Lean proofs must be kernel-checked and then human-audited for statement fidelity.
On October 6, 2026, OpenAI published 722 mathematics manuscripts written by an undisclosed internal model on GitHub, grouped into 372 result categories across 17 fields, with 162 including Lean formalizations; OpenAI said each successful result consumed compute roughly equivalent to three hours of ChatGPT Pro thinking. The claimed results include the four-dimensional Kakeya conjecture, problems related to the Riemann hypothesis and the unique games conjecture, with a 175-page paper claiming every four-dimensional Kakeya set has full Hausdorff dimension 4 and a 97-page paper on the three-dimensional Kakeya maximal conjecture. The work has not been peer reviewed and some manuscripts lack Lean formalization; mathematicians raised concerns about transparency, credit and validation workload, 25 Fields Medal winners issued a statement, and OpenAI withdrew three manuscripts over a sign error and revised 14 others, leaving 719 listed.
Set theorist Asaf Karagila responded to OpenAI's preprint claiming the Partition Principle does not imply the Axiom of Choice, calling it muddled, oddly structured, loosely termed and poorly referenced — the sort of submission that should receive a desk rejection. He said OpenAI released hundreds of incomprehensibly written "solutions" without expert input, effectively shifting the burden of sorting and verification onto mathematicians. He therefore will not send Sam Altman the promised bottle of whisky and will not spend time reading the paper.
On October 6 OpenAI published roughly 722 manuscripts, grouped into 372 "result families," claiming most came from a single internal model; it tested about 8,000 open problems with roughly 5% success and about three hours of ChatGPT Pro-level compute per problem. The Association for Human Mathematicians (AHM), chaired by Fields Medalist Terence Tao, responded that dumping 700-plus files at once is "not a demonstration of scholarship, but a demonstration of power," and urged mathematicians to stop collaborating with OpenAI; Scott Aaronson called it a "Mathocalypse." Three manuscripts have since been retracted, one reportedly for a sign error, with two others depending on it.
OpenAI this week published hundreds of claimed solutions to hard math problems (719 manuscripts, per the article), saying it consulted an advisory group of elite mathematicians to avoid the controversy its earlier result sparked. But only ten of the manuscripts included the model's chain of thought, and by the article's account roughly 42% of the proofs had not gone through formalization in Lean; the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), hosted by Princeton's Institute for Advanced Study, said it is ultimately up to the mathematical community to judge whether its recommendations were followed — its first request being to stop testing advanced problems on proprietary models. A paper from Cambridge and King's College London mathematicians also documents at least two discrepancies between OpenAI's natural-language proof and the Lean code for a problem derived from the Navier-Stokes equations, concluding such autoformalized proofs should not be trusted without peer review.
OpenAI published 722 math manuscripts produced by an unreleased internal model to a GitHub repository, and a 36Kr/机器之心 review picks out the most famous results among them. The manuscripts are grouped into 372 "result families" spanning 17 areas such as number theory, geometry, theoretical computer science and mathematical physics; about 50 families claim falsifications or counterexamples, and roughly 60% include Lean formalization. Claims listed include the quasi-Riemann hypothesis, the BSD formula and Goldfeld's conjecture, Hilbert's tenth problem over the rationals, the Unique Games Conjecture, a Hadwiger conjecture counterexample and the free group factor isomorphism problem — all described as OpenAI's own claims and mostly not yet peer reviewed.
A GitHub repository reports that a complete optimality proof for packing 11 squares has passed Lean verification: an EvolvingPrograms verification run accepted all 7,920 local Lean modules with a final audit reporting zero admissions. The project pins Lean 4.34.1 and a fixed Mathlib revision, and gives the optimal side length as T=(6u+4)/(1+2u−u²), where u is the unique root in (9/25,37/100) of a stated degree-eight polynomial, with the construction attaining about 3.8770835900228141773. Some expensive exact numerical certificate checks use native_decide, so the final theorem trusts Lean's kernel and native compiler rather than claiming kernel-only verification.
A 36Kr report says Anthropic engineers uploaded a Lean-verified proof generated with an unpublished Claude model for a long-standing percolation conjecture in dimensions three through ten. The claim has not been independently established in the source: the original researchers are reportedly waiting for a human-readable explanation, and Anthropic has not published an official announcement.
In an SAIR interview reported by 36Kr, mathematicians Tao and Deng Yu said AI-generated proofs do not necessarily mean humans understand them, and urged model companies to slow batch publication. They proposed earlier mathematician participation, clearer source attribution, formal verification, and a period for the community to interpret results. They also linked their participation in Open Math Model to giving mathematicians a role in tool development, while the team’s new model remained in preparation and Open Math OS was described as the first planned release.
This past year, OpenAI, Anthropic, and other labs have announced breakthroughs on numerous long-standing mathematical problems, in some cases pushing well beyond what researchers expected current systems to be capable of — including resolving one of the famous Millennium Prize problems. But in classic Silicon Valley style, AI labs are moving fast and breaking things, […]
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…