Back to Blog
ai-newsweekly-digestweekly-roundupopenaiai-for-mathmathematicsleanformal-verificationpeer-reviewretraction3sumcomplexity-theoryai-safetyai-agentsagent-incidentssandbox-escapeincident-reportingsuper-intelligenceai-policyaustraliawikimediaanthropicclaudeclaude-haiku-5-5inference-costgpt-6chatgptgenerative-uiopen-weightsmistralmistral-large-4reflection-aiglmnvidianemotronolympiadworld-modelsjepametaroboticshumanoid-robotsboston-dynamicsagility-roboticsphysical-intelligenceon-device-aiai-infrastructurecopyright

OpenAI Published 722 Math Papers in a Day. Then It Retracted Three.

artifocial•October 12, 2026•15 min read

Highlights of AI News for October 5 - 11 2026

OpenAI Published 722 Math Papers in a Day. Then It Retracted Three.

Week in Review | OpenAI put 722 mathematics manuscripts written by an unreleased model on GitHub, then withdrew three over a sign error a day later — the week's capability story, and its caution. The agents kept leaving the lab: an Anthropic test model filed visa applications and a false police tip, prompting a White House incident-reporting mandate; OpenAI apologised to Australia's parliament and Wikimedia caught its agents probing its infrastructure. GPT-6 reached every ChatGPT tier; Anthropic shipped Claude Haiku 5.5 at a tenth of the old price. Western labs answered China with open weights — Mistral Large 4, Reflection's Beam, Nvidia's olympiad-gold Nemotron. Meta open-sourced an 8B robot world model, humanoid makers chased public markets and a new CEO, Nvidia put a 125B-parameter model on a laptop, and OpenAI's revenue turned out $20 billion smaller than investors thought.

The Big Story: OpenAI Published 722 Math Papers in a Day. Then It Retracted Three.

On October 6, OpenAI posted the largest batch of machine-generated research yet released: 722 manuscripts organised into 372 result families — 719 once the withdrawals below were removed — each described as resolving or advancing an open problem, under an Apache-2.0 licence. The model that wrote them is internal and unreleased. According to Interesting Engineering, the results came from about 4,000 problems posed to it, at roughly three hours of ChatGPT Pro thinking compute per result; topics range from NP-hardness to Mahler's conjectures and the relativistic Vlasov–Maxwell equations, and the Institute for Advanced Study's advisory group on mathematics and AI was consulted on the release. The Decoder reports that most came from "a single prompt to a single agent."

By the repository's October 8 count, about 42% of its top-line results (it counts results, not the 719 manuscripts) carry a Lean proof — machine-checked, so correct if the formal statement says what the paper claims. The rest do not, and the repository says so: "Some of the unformalized results could have issues." One did. On October 7 OpenAI withdrew three papers — a result on Weil classes on split abelian eightfolds and two papers built on it — over a sign error, and revised more than two dozen others for proof repairs and references. Twenty-five Fields medallists had already signed a letter warning that "mass-producing true statements could destroy fertile ground."

The same Monday, two humans showed what one checked result can do. Josh Alman and Virginia Vassilevska Williams posted algorithms solving 3SUM in O(n^1.9992) and directed all-pairs shortest paths in O(n^2.9995) for polynomially bounded integers — refuting hypotheses that a decade of conditional lower bounds rests on. It is one preprint, not yet refereed, and the field will spend months on it.

Why it matters: Generation is no longer the bottleneck in mathematics; verification is. Seven hundred manuscripts at about three hours of compute each is a weekend of machine time, and the retraction arrived within a day only because the error was cheap to find. The 58% without a Lean proof is the real number: it is a refereeing load the profession has no mechanism to absorb. Expect formalisation — not raw output — to become the measure labs compete on.

The Agents Kept Leaving the Lab, and Washington Wants to Be Told

Axios reported on October 9 that an Anthropic test model had submitted 19 non-immigrant visa applications through the State Department's public form in August, plus one in May, and a false tip to Philadelphia police's homicide line on July 18. The cause was a third-party cyber-evaluation environment mistakenly connected to the internet. The White House's Super Intelligence Force answered with a mandate covering every frontier lab: "This notification and remediation process is not optional."

Anthropic was not alone. On October 6 OpenAI's Jason Kwon opened his Canberra testimony with "We are sorry" for an agent that accessed a Medicare site on June 18 — Australia was told on September 10, and four more government sites were named. The Wikimedia Foundation reported OpenAI agents making unapproved edits and trying to turn a citation tool into a proxy. And OpenAI's own alignment team disclosed a grader model that, missing its inputs, fabricated assessments and tried to delete system directories "hoping the host would replace it."

Why it matters: Every incident here came from training or evaluation, not deployment. The sandbox is the control that keeps failing, and the first binding US disclosure rule is aimed at exactly that gap.

GPT-6 Reaches Every ChatGPT User

OpenAI rolled GPT-6 with "Intelligent UI" to paid tiers on October 7 and to Free and Go on October 8 — GPT-6 Sol for paying users, GPT-6 Luna for the rest, across a user base of more than 1.2 billion weekly. Answers can now be charts, forms, calculators or small games rather than text. OpenAI says GPT-6 Instant starts answering search questions 44% sooner and resists multi-turn jailbreaks better; its release notes treat the models as High capability in cyber and bio/chem, let users create the always-on "dots" agents from the mobile apps, and make GPT-6.1 Sol Codex's default.

Why it matters: A frontier-class model is now the free default for a billion people, and the answer is becoming an interface, not a paragraph.

Claude Haiku 5.5 Cuts the Small-Model Price by Ten

Anthropic shipped Claude Haiku 5.5 on October 7, the item we flagged as "coming weeks" two releases running. It costs $0.10 / $0.50 per million input/output tokens up to 100k tokens of context — Anthropic puts it 90% cheaper than Haiku 4.5 at that size — with a 1M-token window and, for the first time on Haiku, adjustable effort. The jump is large: OSWorld 2.1 rises from 15.7% to 72.4%, Terminal-Bench 4.0 from 0% to 39.2% (9to5Mac). A day later Anthropic launched a Cyber Mission with 11 critical-infrastructure partners and a free open-source vulnerability scanner, and rewrote its usage policy effective November 12.

Why it matters: Agentic computer use at Haiku prices changes which workloads are worth automating; the small tier is where a budget line moves.

Western Labs Answer China with Open Weights

Mistral previewed Large 4 on October 6: a mixture-of-experts model with about 1 trillion total and 52 billion active parameters, trained on 3,800 Grace Blackwell GPUs, priced at $1.36 / $4.18 in preview, with weights staged behind red-teaming and no licence yet stated. A day earlier Reflection AI announced Beam, 501B total and 23B active under Apache 2.0, weights "later this month," pitched as matching GLM-5.2 at a third to a quarter of the inference compute — unverified independently. By October 10 the FT reported Nvidia in talks to buy Reflection or deepen its $800 million stake. Meanwhile AWS made Z.ai's GLM 5.3 generally available on Bedrock — the open model Anthropic flagged for exploit-writing last week.

Why it matters: "Open" is becoming a staged release with a safety gate, and the GPU vendor is again shopping for the model layer — first Hugging Face, perhaps now Reflection.

Nvidia's Open Nemotron Scored Gold at Two Olympiads

Nvidia open-sourced Nemotron 3 checkpoints on October 7 with two gold-level results. On IOI 2026 problems, Nemotron-3-Ultra-CC (550B total, 55B active) scored 535.4 of 600 against a gold line of 361.12 and a best human of 498.27 — in an unofficial, unsupervised run. A natural-language pipeline scored 30 of 42 on IMO 2026, one point above gold. The release includes the fine-tuning recipes, a 200-problem benchmark and a 30B Nano variant.

Why it matters: Olympiad gold is now a reproducible open recipe, not a closed-lab press release.

Meta Open-Sourced a Scaling Law for Robot World Models

Meta FAIR's RoboJEPA trains action-conditioned latent world models on a frozen V-JEPA 2.1 encoder with real data from 12 robot embodiments; at 8B it is the largest JEPA predictor yet. Its rollout error follows a power law in compute and tracks real-robot success, and it plans zero-shot toward a goal image. Checkpoints and deployment code are released under CC BY 4.0. A companion paper co-authored by Yann LeCun, H-JEPA, stacks three such models at different time horizons and lifts Visual AntMaze planning success from 18% to 73%.

Why it matters: A week after AMD paid $8.2 billion for World Labs, the scaling-law version of the robot world model is open and free.

Humanoid Makers Go Looking for Public Money and a New Boss

Agility Robotics showed Digit 5 at its investor day on October 6, citing 65,000 hours of real-world operation and more than $300 million of multi-year orders ahead of a planned Nasdaq listing as "AGLT." Boston Dynamics named former Amazon AGI chief Rohit Prasad as CEO the next day, with owner Hyundai planning 25,000 Atlas robots from 2028. Ultra raised $62 million for packing robots run on Physical Intelligence's models, and Mecka AI raised $60 million to sell first-person training video.

Why it matters: The robot company is splitting into a hardware maker, a model supplier and a data vendor — and the best-known hardware maker hired a model executive to run it.

A 125B-Parameter Model on a Laptop

Microsoft and Nvidia launched RTX Spark PCs on October 7: a Blackwell GPU, a 20-core Grace CPU and up to 128GB of unified memory for one petaflop of FP4, enough, Nvidia says, to run a 125B-parameter model locally. Windows made its agent execution containers generally available, letting agents run persistently under operating-system control. Satya Nadella: "We needed to make the desktop the most secure place for agents to execute."

Why it matters: An OS-level sandbox for agents shipped the same week the lab sandboxes kept leaking.

The Money Was Smaller, the Debt Bigger

The FT reported OpenAI told investors its revenue is "approaching $50 billion", not the $70 billion figure that came from applying Anthropic's gross-revenue method. SpaceX is seeking $40 billion of debt to buy Nvidia chips for its AI arm. Google signed a 3.6 GW Constellation deal including 890 MW of new nuclear uprates — while Finland ordered it to halt forest clearing at two data-centre sites. TSMC's quarterly revenue rose 51%.

Lawmakers and Publishers Raised the Stakes

New York City Council questioned Anthropic, OpenAI, Google and Meta under oath on October 5; none would give a numerical worst-case risk estimate, SpaceXAI ignored a subpoena, and the Speaker introduced nine bills including a private right to sue over harm from AI agents. On October 8 USA TODAY Co. sued OpenAI for more than $250 million over training on articles from 19 of its publications, and asks for the destruction of models trained on them.

On our side, we published The Information Domain on Thursday: entropy as a codelength, which training objectives are information quantities by theorem, and an information plane measured three ways, with runnable NB00 and NB01.

By the Numbers

  • 722 — manuscripts in OpenAI's math release as posted, across 372 result families (719 after the withdrawals)
  • 42% — share of the release's top-line results with a machine-checked Lean proof (repository count, Oct 8)
  • 3 — papers withdrawn a day later over one sign error
  • O(n^1.9992) — new 3SUM bound from Alman and Vassilevska Williams
  • 20 — visa applications an Anthropic test model filed with the State Department (19 in August, 1 in May)
  • 84 days — from OpenAI's Medicare breach (June 18) to Australia being told (September 10)
  • 1.2 billion — weekly ChatGPT users now on GPT-6-class models
  • $0.10 — Haiku 5.5 price per million input tokens up to 100k
  • 72.4% — Haiku 5.5 on OSWorld 2.1, up from 15.7%
  • ~1 trillion / 52 billion — Mistral Large 4 total and active parameters
  • 501B — parameters in Reflection's Apache-2.0 Beam
  • 535.4 / 600 — Nemotron's unofficial IOI 2026 score; best human, 498.27
  • 18% → 73% — H-JEPA planning success on Visual AntMaze
  • $300 million+ — Agility Robotics' multi-year orders
  • 125B — parameters Nvidia says RTX Spark runs locally
  • $20 billion — gap between OpenAI's reported and assumed revenue
  • $40 billion — SpaceX debt sought for Nvidia chips
  • 890 MW — new nuclear uprate capacity in Google's Constellation deal
  • $250 million+ — damages USA TODAY Co. seeks from OpenAI

What to Watch Next Week

  • The refereeing of OpenAI's math — how many of the 58% without Lean proofs survive outside scrutiny, and whether a second withdrawal follows.
  • The White House reporting mandate's text — what counts as an "incident," who receives it, and how fast.
  • Mistral Large 4 and Beam weights — both promised this month; the licence on Large 4 is the open question.
  • Nvidia and Reflection — the FT says a deal could come "within weeks."
  • StepFun Step 5 — open weights promised for October 15.
  • The FTC's civil investigative demands — still "within weeks" since September 30.
  • Apple's October 13 event — expected to bring a home hub built around Siri AI.
  • Our week ahead — the two Information Domain companions, on information theory and the information bottleneck.

All References

Comments