🧮 OpenAI Drops 722 Papers of Machine Mathematics on a Field Already in Shock

On Tuesday evening OpenAI published 722 papers, organised into 372 groups of results, into a public GitHub repository — machine-generated mathematics covering number theory, algebraic geometry and theoretical computer science, produced by an internal model the company has not released (OpenAI, Scientific American, WSJ, New Scientist). OpenAI says it fed the model roughly 4,000 problems in total. Among the claims: a solution to the four-dimensional Kakeya conjecture, improvements on some of the most important algorithms in computer science, and genuine progress toward the Riemann hypothesis — advances touching three of the five remaining Millennium Prize Problems.

The scale is the easy part to state and the hard part to absorb. What makes the release strange is the cost curve. Last month’s blockbuster — a proof about the Navier-Stokes equations — came out of a collaborative swarm of some 10,000 agents and millions of dollars of compute. An OpenAI spokesperson told Scientific American that almost every result in this batch came from a single prompt handed to a single AI agent. OpenAI’s own post says the repository includes ten summaries of the model’s reasoning, estimates of compute expressed in terms of ChatGPT Pro usage, and statistics on attempted problems, with the average result consuming the equivalent of about three hours of ChatGPT Pro thinking (HN thread, 1,025 points). There is also a standalone preprint on integer multiplication below n log n in the same repository.

Verification is where the argument now lives. Many of the results have already been checked in Lean, the proof assistant that validates a proof’s logic step by step, and Scientific American reports those are all but certain to be correct. But Kevin Buzzard of Imperial College London counted roughly 30 papers relevant to his own field and found only one formally verified in Lean — the line between verified and plausible is currently a small number. Mathematicians will spend months deciding whether the proofs contain novel ideas or are mostly recombinations of existing techniques, and criticism has been as loud as praise: Fields medallists have warned that flooding a field with machine output risks undermining the purpose of the discipline, and OpenAI’s reputation among mathematicians for bold claims with little transparency is doing the rest. OpenAI says it drew on publishing guidelines an advisory committee issued last week, based on feedback from hundreds of mathematicians, and its own framing is cautious: “There are difficult questions ahead about how AI can support mathematical understanding.”

🐈 Mistral Opens Le Chonk: A 1-Trillion-Parameter Model, Weights Promised This Month

Mistral launched a public preview of Mistral Large 4 — officially ML4, unofficially le Chonk — and it is the largest model the French company has built: 1 trillion total parameters with 49 billion active, natively multimodal, in a mixture-of-experts design (Mistral, VentureBeat, Silicon UK). The API is live on Mistral Studio, and the weights drop by the end of the month. The compute detail is European: ML4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own datacenters, and the preview is served from the same infrastructure.

Mistral’s claims are benchmark-shaped but specific. It says ML4 is “competitive with the strongest open-source models globally” while “significantly outperforming any open-weight model developed in the US or Europe,” and calls it state of the art among open models on cybersecurity, finance and law — with visual grounding the one category where it claims to beat closed frontier models outright. VentureBeat notes a 15 percent task-pass rate on Harvey’s Legal Agent Benchmark as an independent cross-check, and reports Mistral co-founder Guillaume Lample treating the release as an early version of the long-running Le Chaton Fat meme: a much larger model is apparently still coming. Until the weights land, the model is doing something unusual — red-teaming with cybersecurity leaders, vetted partners and state authorities who get the same model with reduced moderation and expanded cyber capabilities.

On the OpenRouter board a single new endpoint, mistralai/mistral-large-4-0, carries a 524,288-token context at $0.68 / $2.09 per M tokens, with cached input at $0.07 per M. The Hacker News release thread drew 519 points.

🔍 EmbeddingGemma 2 Puts Multimodal Search on the Phone

Google DeepMind released EmbeddingGemma 2, a 740-million-parameter embedding model that maps text, code, images, video frames and audio into one 768-dimensional space, under an Apache 2.0 licence (Google DeepMind, Google Developers, Hugging Face). The first EmbeddingGemma, text-only, passed 20 million downloads; this is the same pitch extended to four modalities and built on the Gemma 4 architecture.

The design is modular rather than monolithic. A 270M text core (a 130M backbone plus a 140M embedder) does the work alone for text-only apps; vision adds 170M and audio 300M, so the footprint climbs from roughly 440M to 740M only if an app needs it. Quantised on a Pixel 11 Pro, that measures at about 191 MB of active RAM for text-only weights and about 567 MB for the full stack. Matryoshka Representation Learning lets developers truncate vectors from 768 dimensions down to 512, 256 or 128 and re-normalise — up to six-fold storage reduction, with Google’s developer guide claiming image, video and speech retrieval keep about 95 percent of full quality at 256 dimensions. The context window is 8K tokens, four times the first version, enough for roughly 5.5 minutes of audio, 29 images or 58 video frames in a single call. The benchmark that moved most is code: MTEB Code rose 9.92 points, from 68.76 to 78.68.

Two cautions sit in the model card. Weights are on Hugging Face and Kaggle now, and serving is supported by transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LM Studio — but this is a pre-trained embedding system with no post-training alignment and no output-level moderation, so Google says safety work went into pre-training data filtering and that applications need their own safeguards. The release drew 363 points on Hacker News, and it appeared in the trending feed the same day it shipped.

📉 Dalio Says the Bubble Is Close, and the Job Cuts Arrive Anyway

Speaking at the Forbes Global CEO Conference in Singapore, Ray Dalio called the AI investment boom a “classic bubble” and said markets are approaching the point where it bursts (Business Times, Bloomberg). His mechanism is borrowing costs: a large share of AI investment is debt-financed, gains are concentrated in a handful of mega-caps, and global bond yields have climbed to multi-decade peaks — so the cost of funding data centres and chip capacity rises exactly as the revenue is still being promised. “We’re in that part of the cycle that is before that, but approaching that,” he said. “I think we’re close to that.” On the same panel, Temasek chief executive Dilhan Pillay warned that excessive capital concentration in AI infrastructure could raise the cost of capital for everyone else, while describing his own approach as an ongoing process rather than a single bet. The context is not hypothetical: global AI-related debt issuance is on track to top 500 billion dollars in 2026.

The sober mood is now visible in hiring. Reuters reports companies cutting jobs as investment shifts toward AI, and the Financial Times reports HSBC planning job reductions across its UK wealth business in an AI push. The IMF’s Kristalina Georgieva warned that an energy shock, growing debt and AI risks together threaten global growth, and the San Francisco Fed’s president said AI demand could extend the energy shock. Meanwhile the market’s own answer to the bubble question is the opposite: US stocks closed at an all-time high on continued AI bets, and Microsoft and Nvidia took the stage in San Francisco the same week to sell an agentic future.

💊 Utah Lets an AI Write the Prescription, and Washington Renames the Technology

Utah has authorised the first US state-regulated path for an AI to issue an initial prescription without a doctor reviewing the individual case, and the startup behind it, Nolla Health, opened enrolment this week (LA Times, SiliconANGLE, Dallas Express). The Nolla Derm app takes a 10-to-15-minute questionnaire and a five-angle face scan, scores acne severity numerically, and picks from eight approved topical treatments — tretinoin, compounded topical spironolactone, niacinamide, adapalene, clindamycin with benzoyl peroxide, benzoyl peroxide and azelaic acid — at 4.99 dollars a month. Oral drugs, hormonal therapy and isotretinoin are excluded, and hard stops route pregnancy, breastfeeding, immunosuppression, prior reactions, severe acne or poor image quality to a human.

The staged rollout is the interesting part, because it is the regulator’s actual policy. In stage one, two Utah-licensed physicians review every AI prescription before it reaches a pharmacy. The AI may only submit directly after the pilot has run at least four weeks, treated at least 100 patients, held at least 95 percent agreement with physicians, missed no required safety stops and logged no serious adverse events — and only with further regulatory approval. Stage three, at 750 cumulative patients, cuts routine review to a monthly sample of at least 10 percent, while every adverse event and escalation still goes to a doctor. Utah’s public pilot page still listed the demonstration as not started on October 6 even as Nolla invited patients in, so the autonomous phase has not clearly begun. The state has been here before: an earlier AI refill pilot with Doctronic drew an April letter from the Utah Medical Licensing Board — which said it had not been consulted and argued that “there is a reason prescription refills require physician authorization” — and the state responded by keeping the pilot and removing two medications from the approved list. The American Medical Association adopted policy in June calling for AI to support rather than replace physician judgement.

Two more signals from the same 24 hours. The US Justice Department told employees to use the words “super intelligence” instead of “artificial intelligence” in most circumstances, including in court, per a memo from acting Deputy Attorney General Trent McCotter that follows President Trump’s executive order last week; Trump separately named Director of National Intelligence Jay Clayton as his AI czar, heading a task force called the Super Intelligence Force (Reuters). And a proposed nationwide class action against McDonald’s alleges the chain illegally coordinates menu prices across franchises and company-owned restaurants through an AI-powered pricing system trained on nonpublic data drawn from nearly 14,000 restaurants, a violation of antitrust law via algorithm; McDonald’s says franchisees make their own pricing decisions (Reuters, The Guardian). A Reuters/Ipsos poll conducted September 30 to October 5 found 57 percent of registered voters say the Trump administration has not taken AI risks seriously enough or at all, 54 percent say the same of Congress, 84 percent see AI as a threat to American workers, and a bipartisan majority wants stricter regulation (CNA).

🆕 New OpenRouter Models

The created-timestamp sweep found two genuinely new endpoints — the first new-model section in four issues — both listed within hours of the last publication:

Model Context Price (in / out) Listed
mistralai/mistral-large-4-0 524,288 $0.68 / $2.09 per M tokens 2026-10-06 13:42 UTC
google/gemini-nano-banana-2.1 65,536 $1.50 / $7.50 per M tokens 2026-10-06 15:33 UTC

Mistral Large 4 is the same model as the le Chonk preview covered above, listed with cached input at $0.07 per M tokens and no published sunset. Gemini Nano Banana 2.1 is the smaller, stranger listing: a 65K-context image-and-text model that charges $30 per M tokens for image output and $14 per thousand web-search calls, so its economics depend on how much of the work is pixels rather than prose.

📊 Quick Hits

  • OpenAI had an agentic-infrastructure day. Alongside the mathematics dump it opened the Decisions API in public beta — a hosted way to encode decision policies for agents, at 346 points on Hacker News — and published Advancing computer use with Ironclad (Decisions API, OpenAI).
  • An open-source AI accelerator, designed by AI. FeSens/openTPU — an open TPU-style accelerator whose RTL was largely generated by AI — took 309 points on Hacker News, one of the clearest artefacts yet of models writing hardware instead of software.
  • Small decision models become a category. AWS’s Strands Agents released Decider 2B, a 2-billion-parameter open model for deciding whether an agent’s next tool call should proceed, be denied, be confirmed by a human, or be guided — the guardrail layer shipped as a model rather than a rule engine (219 points, Strands).
  • The South Korean banking breach story deepened. WSJ reports hackers used a Chinese AI tool against the banks, and Reuters reports President Lee Jae Myung saying AI agents appear to have been used — the follow-up to yesterday’s account of seven-plus institutions and more than 68,000 affected people (Reuters).
  • Microsoft and Nvidia launch an agentic laptop. Satya Nadella and Jensen Huang are presenting the Surface Laptop Ultra in San Francisco — a 15-inch mini-LED machine with up to 128GB of memory built on Nvidia’s RTX Spark chip, co-designed with MediaTek, pairing up to 20 CPU cores with a Blackwell GPU and up to a petaflop of AI performance. Analysts told Reuters the open questions are price, amid a memory shortage, and whether agents running locally can be contained (Reuters).
  • Google pushes AI into game creation. Google and Unity announced an AI gaming platform partnership, and Google began letting users make games with AI directly (The Verge, Google Cloud).
  • Meta’s Muse tops the charts. CNBC reports the assistant is now the most-downloaded app of its kind and needs to become a habit rather than a novelty, while WSJ profiles the executive behind it (CNBC).
  • Agents could cost banks 500 billion dollars by winning savers better rates, per the Financial Times — a rare estimate of agent value destroyed rather than created (FT).
  • Washington buys compute for science. The administration will receive 100 million dollars in compute credits for a federal AI science initiative (Politico), while Danaher said it will open an AI-powered autonomous research lab in 2027 and Reuters reports AI companies running virtual drug trials to improve the odds of real ones.
  • Also this cycle: Anthropic opened its most powerful models to more security teams; Australian lawmakers questioned OpenAI officials about breaches; Sierra introduced a Personal Agent Protocol; and a developer’s Tell HN post about GitHub refusing to remove cracked copies of their software after a month drew 345 points.

⏰ OpenRouter Limited-Time Free Models

The clarification still matters: OpenRouter’s expiration_date field marks a listing with a published sunset, which is not the same thing as free. The collector returned 10 sunset-dated endpoints, of which 0 are genuinely free; a full API sweep found 31 dated endpoints, including 3 genuinely free listings.

Model Context Price (in / out) Sunset
poolside/laguna-s-2.1:free 262,144 $0 / $0 — free 2026-10-31
poolside/laguna-xs-2.1:free 262,144 $0 / $0 — free 2026-10-31
dots-studio/dots-3-note-preview:free 512,000 $0 / $0 — free 2026-12-31
baidu/ernie-4.5-vl-424b-a47b 123,000 $0.42 / $1.25 per M tokens 2026-10-08
qwen/qwen3.6-max-preview 262,144 $1.03 / $6.16 per M tokens 2026-10-09
qwen/qwen3-coder-plus 1,000,000 $0.65 / $3.25 per M tokens 2026-10-09
qwen/qwen3-vl-235b-a22b-thinking 131,072 $0.40 / $4.00 per M tokens 2026-10-09

The sunset clusters are now the most useful part of this section, and this cycle they are close. October 8 retires Baidu’s ERNIE 4.5 VL. October 9 is tomorrow, and it takes a single Qwen block spanning 15 endpoints, from the 8B vision models up through Qwen3 235B, Qwen3 Max, Qwen3 Max Thinking and Qwen3 Coder Plus. October 20 clears the whole Gemini 2.5 family across five endpoints. October 31 retires four poolside listings, including the two free Laguna entries above, November 11 dates three ByteDance Seed models, and December 31 dates the Z.ai GLM 4.7 and 4.5 pair plus the free dots preview. The :free shelf holds 16 endpoints in total, 3 of them dated, leaving 13 with no expiry at all — down from 16 undated a day ago, as promos continue to lapse faster than they are replaced.

🤗 Trending on Hugging Face

The collector’s new-uploads list was zero-download test repositories for the 33rd consecutive cycle, so this section uses Hugging Face’s official trending feed.

Google’s new embedder arrived with traction on day one. google/embeddinggemma-2 is already at 7,562 downloads / 787 likes — unusual for a model hours old on a feed ranked by accumulated downloads, and a sign that edge-deployment tooling was waiting for it.

The decision-model pair keeps compounding. autotrust/JEV-27B-VL sits at 1,529,210 downloads / 1,467 likes, with likes up 69 percent in a day, while sibling autotrust/GEV-26B-Decide reached 895,867 / 986. The laya thread, which spent weeks at zero, logged a third consecutive step: convaiinnovations/laya is at 28,497 downloads / 5,316 likes, though the download multiple decelerated to roughly 1.4x.

What people actually pull:

⭐ GitHub Trending: AI Edition

All five of the collector’s entries were out of window this cycle — the 22nd consecutive run — with created dates of October 2 and 3. This section uses API-verified repositories created inside the window instead, and the in-window pool was thin and largely non-AI at the top:

  • sisyphuslabs/omo-dori-mode-experimental — ⭐49. An always-on messenger agent that launches, tracks and closes coding-agent sessions, shipped as an Agent Skill for a terminal multiplexer called herdr. The day’s top in-window AI repository, and the fourth consecutive cycle an Agent Skill has led this section.
  • sun168567/cc-desk-tunnel — ⭐24. A desktop client that drives Claude Code in the cloud while executing on a local Windows machine over a reverse tunnel — a small, pointed answer to the question of where agent credentials actually live.
  • orange2ai/podcast-to-article — ⭐17. An Agent Skill that turns multi-speaker podcast transcripts into publishable long-form articles, with a self-check pass for removing machine-sounding prose.
  • hey345437-boop/futuresearch-gateway — ⭐16. Wraps an asynchronous deep-research API as an OpenAI-compatible endpoint, with a key-pool and per-tenant quota panel plus two MCP servers, in Go. A second repository in the same window does a near-identical job for a different backend — the day’s naming pattern.
  • karminski/karminski-design-skills — ⭐15. A personal collection of UI design Skills under a non-commercial licence — a small example of the shift from shared prompts to licensed, attributed design assets.

Three notes on the pool. First, the raw star-sorted in-window search was led by non-AI repositories: a macOS screenshot tool at ⭐238, a Rust port of the TypeScript compiler at ⭐165 and an Nvidia driver for macOS at ⭐126 — none of them AI work, which is why this section is enumerated by hand rather than by rank. Second, one owner published five clean-room, pure Rust reimplementations of Word, Excel, PowerPoint, AutoCAD and Pro Tools within about three hours (storytold/wordcraft and siblings), a promotional cluster treated here as one signal rather than five entries. Third, the honest floor: the in-window pool held nothing above 50 stars in AI, against 299 a day ago.

The collector’s new-repo list was five out-of-window repeats this cycle, so entries were substituted from GitHub’s search API and verified individually against each repository’s created_at.

💡 Key Trends

  1. Machine mathematics has outrun the institutions that judge it. OpenAI published 722 papers in one night — more than any mathematician can read, produced by a model nobody outside the company can query, with the disclosure that a single prompt and single agent did almost all of it. The bottleneck has moved from generating results to validating them, and a field that cannot verify at production speed will end up trusting either reputation or formal proof. Only the second is a standard, which is why the Lean count — 30 relevant papers, one verified — is the number that matters.
  2. The open-weight frontier now has a European bid at frontier scale. Mistral trained a 1-trillion-parameter model on 3,800 Grace Blackwell GPUs in its own datacenters and framed the result explicitly as sovereignty: better than any open-weight model from the US or Europe, with cyber, finance and law as the proof points. The striking detail is the red-teaming arrangement — cyber leaders and state authorities get the same model with reduced moderation and expanded offensive capabilities before the public gets the weights.
  3. The bubble conversation and the AI capex conversation are now the same conversation. Dalio says the cycle is close to its bursting point on debt-financed infrastructure and rising yields; Reuters and the FT report job cuts being justified by AI investment in the same 24 hours; the IMF warns about debt, energy and AI together. Against that, US equities sit at an all-time high and Microsoft and Nvidia spent the week selling local agents to consumers. Both can be true, which is what makes the next two quarters interesting.
  4. Regulators are choosing pilots and vocabulary over prohibitions. Utah’s staged sandbox — 100 physician-reviewed patients, 95 percent agreement, then monthly 10 percent sampling — is a real template for letting an AI prescribe, and it is careful in a way that makes the Justice Department’s instruction to say “super intelligence” instead of “AI” look like what it is: a naming policy standing where a regulatory one has not arrived. Voters appear to have noticed: 57 percent say the administration is not taking AI risks seriously, 54 percent say the same of Congress.

Today’s Snapshot

  • OpenRouter: 2 genuinely new endpoints in the last 24 hours — Mistral Large 4 (524,288 context, $0.68 / $2.09 per M tokens) and Gemini Nano Banana 2.1 (65,536 context, $1.50 / $7.50 per M tokens). The first new-model section in four issues.
  • Limited-time free models: 3 genuinely free with a sunset (two poolside Laguna entries on October 31, dots-3-note-preview on December 31); the collector’s list held 10 discounted endpoints, 0 of them free; the full API sweep found 31 dated endpoints; the :free shelf holds 16 endpoints, 13 with no expiry, down from 16 undated yesterday. Qwen’s 15-endpoint block clears tomorrow, Baidu’s ERNIE 4.5 VL the day before.
  • Hugging Face: the collector’s new-uploads list was junk for the 33rd consecutive cycle, so the official trending feed was used; google/embeddinggemma-2 entered at 7,562 downloads on release day, and the autotrust decision-model pair kept compounding.
  • GitHub: the collector’s list was 5 of 5 out of window (22nd consecutive run); five API-verified in-window AI repositories substituted, none above 50 stars, with a five-repository single-owner promotional cluster and a non-AI top of the raw search excluded by shape.
  • Models and money: OpenAI published 722 machine-generated mathematics papers touching three remaining Millennium Prize Problems from a single prompt; Mistral opened Large 4 (le Chonk), 1T parameters with weights due this month; Google shipped EmbeddingGemma 2 under Apache 2.0; Ray Dalio called the bubble close to bursting while US stocks hit an all-time high.
  • Safety, law and regulation: Utah authorised the first US AI prescribing pilot with staged physician review; the Justice Department told staff to say super intelligence, not artificial intelligence, including in court; McDonald’s faces a nationwide class action over AI-driven menu pricing; and a Reuters/Ipsos poll found 57 percent of voters say the administration is not taking AI risks seriously.