Eight signals in three days: 372 families of mathematics, an interface built for each question, and a watermark checker open to anyone

In the three days to October 8, 2026, eight items from eleven of the fourteen AI and tech accounts this channel follows were worth carrying, and one subject ran through most of the window: what a frontier model does when it is pointed at open research problems. OpenAI published 372 families of new mathematical results produced by an unreleased model and handed them to a community that has only started to read them. Two labs shipped cheaper models within minutes of each other, Google put its watermark detector in anyone's hands, and Paul Graham argued that a platform which bans buying agents has just marked where the demand is. The window runs three days because the editions of October 6 and October 7 were not published.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

OpenAI published 372 families of new mathematics results from an unreleased model OpenAI's research page and its GitHub repository, with posts from Greg Brockman and Sam Altman · 38,403 likes and 20.0 million views on OpenAI's announcement at capture · October 6

On October 6 OpenAI released a repository of mathematical manuscripts produced by an internal frontier model, described as a broad range of new results across open problems 1. The repository's own catalogue lists 719 manuscripts organised into 372 families, where a family bundles a principal result with its companion arguments, consequences or alternative proofs, and about 42% of the top-line results carry a Lean formalisation — a proof written in the Lean programming language, which lets a computer check it 2. OpenAI says the model was posed roughly 4,000 problems over the evaluation, and that each published result used about three hours of ChatGPT Pro-level thinking compute on average [cite:2].

The list of what came out gives the scale some shape. It includes the Unique Games Conjecture, a statement in complexity theory about how hard it is to approximate a family of optimisation problems even slightly better than a standard relaxation achieves, and L=BPL, a derandomisation conjecture that says randomised and deterministic logarithmic-space computation are the same thing 3. P≠NP is absent from the release, and so are P=BPP and NEXP⊄P/poly [cite:3].

OpenAI says it drew on the advice of the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, an independent body whose members include the mathematicians Timothy Gowers, Edward Witten, Martin Hairer and Melanie Matchett Wood 4. The group's own statement, published the same day, says its advisory role "should not be interpreted as a judgment of the impact of these results or an endorsement of the process by which OpenAI obtained them", and that only the mathematical community can carry out the assessment that is needed [cite:4]. The repository drew 1,280 points and 1,453 comments on Hacker News 5, and 38,403 likes and 20.0 million views on X 6. Sam Altman posted the next morning that the company was "entering a new era of discovery" 7.

The mathematicians who opened the papers needed a model's help to read them Scott Aaronson's Shtetl-Optimized blog, carried by Ethan Mollick, with posts from François Chollet and Yann LeCun · 886 likes and 84,067 views on Mollick's post at capture · October 7–8

Scott Aaronson, a theoretical computer scientist at the University of Texas at Austin, wrote the day after the release that the Unique Games Conjecture is among the results, and that his wife, the complexity theorist Dana Moshkovitz, had worked toward it for the whole time he has known her 3. His account of what the manuscripts are actually like is the part worth reading: as of that evening, humans appear to have understood almost none of the proofs, and Moshkovitz, reading one of them, found "some alien craziness" — a new construction in place of the two objects a proof of that type was expected to use [cite:3]. The paper, in her description, is hard enough to follow that she needed a model to reconstruct its completeness and soundness claims [cite:3]. Ethan Mollick, who teaches at Wharton and writes about how AI is adopted, circulated the blog post as "early first-hand accounts of encountering a narrow superhuman intelligence" 8, and drew the wider conclusion himself: everything ever published will be re-read and re-judged, and novel discoveries will start to come fast [cite:8].

François Chollet, who created the Keras deep-learning library and co-founded the ARC-AGI benchmark, used the same week to ask whether any of this carries. His question is whether mathematics and code can be pushed arbitrarily far with training on verifiable rewards while everything else plateaus on human-generated data, and whether a general ability emerges from that push or only sharper skill inside the two domains 9. His answer, posted the next day: general intelligence looks like a property of the brain rather than a skill that training installs, because the 1980s research programme on the "cognitive consequences of programming" found that learning formal logic through code made students good at code and left their general reasoning where it was 10.

Yann LeCun, a Turing Award winner, answered in French the same morning that mathematics is entering a new era in which formal proof is largely automated and the weight moves to new concepts, abstractions, definitions and conjectures — "the invention of the boat reduced the importance of swimming, but allowed the discovery of new lands" 11.

ChatGPT can now build its own interface for the question you asked OpenAI's release page, with posts from Sam Altman, Greg Brockman and Ethan Mollick · 19,701 likes and 4.0 million views on OpenAI's post at capture · October 7

On October 7 OpenAI began rolling GPT-6 out to all ChatGPT users with a capability it calls Intelligent UI: the model composes a response out of text, visuals, charts, buttons, forms and interactive elements, and decides the mix from the question 12. The engineering behind it is a library of native, streamable components plus a compiler that renders the interface while the model is still generating it, so the layout appears progressively instead of at the end of the answer [cite:12]. OpenAI puts weekly ChatGPT use at more than 1.2 billion people [cite:12].

Two practical details decide whether a reader sees this today. The rollout went to Plus, Pro, Business and Enterprise subscriptions first, and expanded to Free and Go the next day; paid tiers run on GPT-6 Sol and the free tiers on GPT-6 Luna, both tuned for everyday conversation, and the change covers the Chat tab only — the models behind Work and Codex are untouched [cite:12]. On speed, OpenAI says GPT-6 can begin answering while it is still thinking, and that on questions needing a web search GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant; in an internal comparison, GPT-6 Extra High begins answering in the same time as GPT-5.6 Medium while scoring better than GPT-5.6 Extra High [cite:12].

Sam Altman's post announcing it — "ChatGPT can now generate a custom UI for you!" — drew 8,836 likes and 1.0 million views 13, and Greg Brockman's drew 3,456 14. Ethan Mollick, who had early access, reported that it "was a nice change from walls of text" and predicted interfaces built on demand for the problem at hand 15. On Hacker News the release drew 635 points and 348 comments 16.

Claude Haiku 5.5 undercuts its predecessor by 90%, and Anthropic widens the door for security work Anthropic's release page and programme announcement, Simon Willison's weblog · 6,503 likes and 321,059 views on Anthropic's post at capture · October 6–7

Anthropic released Claude Haiku 5.5 on October 7, calling it the cheapest, fastest and most capable small model it has shipped and pricing it for high-volume work such as summarisation, classification and subagent tasks 17. For prompts up to 100,000 tokens — the band holding about 90% of requests to the previous Haiku — it costs $0.10 per million input tokens and $0.50 per million output tokens, 90% below Haiku 4.5, and 50% below the old model above that threshold [cite:17]. The benchmark gaps over Haiku 4.5 are wide: 1620 against 735 on the GDPval-AA v2.1 knowledge-work evaluation, 72.4% against 15.7% on OSWorld 2.1 computer use, and 45.9% against 10.2% on Humanity's Last Exam without tools [cite:17].

Simon Willison added the caveat a price list hides: Haiku 5.5 exactly matches the price of OpenAI's GPT-6 Luna up to 100,000 tokens, but it ships a less generous tokenizer that turns the same long prompt into about 1.25 times as many tokens, and above 100,000 tokens Luna becomes the cheaper option 18. The same day Anthropic halved the price of cache reads on Sonnet 5.5 and started giving Max and Team subscribers monthly API credits — $100 for Max 5x, $200 for Max 20x and up to $500 pooled for Team — which expire at the end of each month [cite:17][cite:18]. The release drew 879 points and 423 comments on Hacker News 19.

A day earlier Anthropic had widened access from the other direction. Its Cyber Verification Program now has three tiers — Defense, Red Team and Specialized — each granting vetted security teams access to the company's most capable models with cyber safeguards reduced to match the work, from incident response and malware analysis up to authorised penetration testing and the testing of systems such as power grids and interbank transfer networks 20. Anthropic reports that partners in its earlier Project Glasswing programme uncovered at least 129,000 verified software vulnerabilities between April and July 2026, with more than 33,000 rated critical or high severity so far, and says the real figure may be at least five times higher because fewer than half of the partners disclosed their patch numbers [cite:20]. The announcement drew 5,706 likes and 1.2 million views 21.

Mistral's 1-trillion-parameter Large 4 wins the vulnerability test that closed models refuse Mistral's release post and the announcement Simon Willison quoted · 755 likes and 30,619 views on Willison's post at capture · October 6

Mistral opened a public preview of Mistral Large 4 on October 6, a natively multimodal mixture-of-experts model with 1 trillion parameters of which 52 billion are active on each token, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company's own European datacentres 22. The weights are promised by the end of the month, and Mistral is positioning the model squarely on sovereignty: a European deployment it operates end to end, training data spanning more than 160 languages including every official European Union language, and a €3 billion Series D behind the compute build-out [cite:22].

Its sharpest claim is about what it will do rather than how well it scores. On the Artificial Analysis Cyber Index, an independent evaluation of finding and fixing security flaws in real software, Mistral says ML4 sits among the top five models globally and leads open-weight models built outside China; on one test that asks a model to reproduce a real open-source vulnerability and then patch it, ML4 scores 82%, the highest of any model [cite:22]. Mistral attributes that to a practical difference: several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same test because they refuse the task, and defending software often begins with proving a flaw is real [cite:22]. The preview is available on Mistral's API at $1.36 per million input tokens and $4.18 per million output tokens [cite:22]. Hacker News made it the largest single discussion in this window, at 2,013 points and 1,197 comments 23. Simon Willison's own test, posted as "Mistral can pelican now!", drew 755 likes 24.

An open 740M model puts image, audio and video search on the phone Google's developer blog and Google DeepMind's thread on X · 4,356 likes and 789,092 views at capture · October 6

An embedding model turns content into numeric vectors so that similar things land near each other, which is the machinery under search and retrieval over a private set of files. Google's EmbeddingGemma 2, released on October 6 under the Apache 2.0 licence, does that for code, images, audio and video in one shared space from 740 million parameters, built on the same Gemma 4 architecture and technology as Google's Gemini embedding models 25. Against its predecessor it improves code retrieval by 9.92 points on the Massive Text Embedding Benchmark's code track, from 68.76 to 78.68, and its 8,000-token context is four times the first version's, enough for 5.5 minutes of audio, 29 images or 58 video frames in one pass [cite:25].

The numbers that make it a phone model are memory ones. Matryoshka representation learning lets developers truncate the output vector from 768 dimensions down to 128 for up to six times less storage; on a Pixel 11 Pro the text-only weights need as little as 191MB of active memory and the full multimodal model about 567MB, and a text-only workload can run on 270 million parameters with vision and audio encoders of 170 million and 300 million added as needed [cite:25]. The first EmbeddingGemma passed 20 million downloads, and the new weights are on Hugging Face and Kaggle [cite:25]. Google DeepMind's thread drew 4,356 likes and 789,092 views 26, and the release drew 422 points and 46 comments on Hacker News 27.

Anyone can now test a watermark, while text stays the hard case Google's announcement, OpenAI's policy page, and both companies' posts on X · 10,893 likes and 7.9 million views on OpenAI's post, 1,291 likes on Google's, at capture · October 5–7

Google opened its SynthID Detector to everyone on October 7, available globally in English, so that any reader can check whether an image, video or audio file was made with Google AI or with tools from partners including OpenAI, NVIDIA and Kakao, with Apple joining soon 28. Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio, and that the built-in verification it already runs in Search, the Gemini app and Chrome handles over a million requests a day [cite:28]. Google DeepMind's post announcing the change drew 1,291 likes 29.

Two days earlier OpenAI laid out the same problem for text, where the question is harder and the answer is deliberately partial. From the coming weeks it will add an invisible watermark, textGrain, to eligible ChatGPT and Codex output inside the European Union to satisfy the EU AI Act, while API customers anywhere can opt in now and the detector goes only to approved researchers 30. OpenAI published the measurements behind that caution: at a target false-positive rate of 1%, the detector found watermarks in about 80% of 200-token passages and about 95% of 400-token ones, with mathematics much lower because the wording has less room to move; replacing 10% of a passage's words with synonyms dropped detection from about 92% to 66%, and replacing 25% dropped it to 17% [cite:30]. OpenAI reports no meaningful benchmark differences between watermarked and unwatermarked output [cite:30]. Its announcement post drew 10,893 likes and 7.9 million views 31.

A platform that bans buying agents has shown a startup where the demand is Paul Graham's and Naval's posts on X, with GeekWire's report on Amazon · 6,031 likes and 1.6 million views on Graham's post at capture · October 7–8

Amazon cut off Meta's Muse personal agent from shopping on Amazon.com in September, showing users who tried it a message that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use", after asking Meta to remove Amazon from the experience 32. Amazon's stated objection is that the agent browsed without identifying itself and appeared to capture and store customer credentials, and the stakes on its side are commercial as well as security ones: it generated more than $68 billion in advertising revenue the previous year, a business that depends on people browsing its pages and seeing sponsored products [cite:32].

Paul Graham read the ban as a market signal, posting on October 8 that "any business that bans agents represents an opportunity for a startup to create a competitor. If they need to ban agents, it must mean people want to use them, or there would be no point. And that in turn means there's demand for a competitor that allows them." 33 His first post drew 6,031 likes and 1.6 million views, and two follow-ups drew 1,318 and 1,138 likes 34.

Naval reached the same question from the other side the day before, arguing that models are the last moat left in software: a model can train on published essays and rewrite them, decompile a program and recode it, or ingest art and recreate it, but a model itself resists being distilled — so more software will retreat to the server where its weights cannot be lifted 35. He was quoting a claim that Photoshop had been reimplemented as open source by decompiling it into a specification and generating Rust from that [cite:35]. His post drew 5,646 likes and 839,584 views [cite:35].