Wednesday's Rundown led with OpenAI saying an unreleased internal model solved Navier-Stokes — one of the seven million-dollar Millennium Prize maths problems — while two outside mathematicians disputed how the race played out. I rewrote it the way I would tell a colleague over coffee. Short version: OpenAI says about 10,000 agents ran for 88 hours on a model "significantly more capable" than GPT-6 Astra, at a compute cost in the "millions of dollars." NYU's Tristan Buckmaster and Anthropic's Levent Alpöge say they had spent a year on a similar path and posted partial results the night before. Meta launched Muse, an always-on personal agent with its own cloud computer. OpenAI shipped ChatGPT Images 2.5. A how-to covers checking where your brand shows up in AI search answers. Reader Sam Reti built an archery scorer from 3,000 hand-labeled photos.
The newsletter also carried ads for Optimizely, LangChain, and a Tempo report. Skipping those.
Words worth knowing
| Word | In one line |
|---|---|
| Navier-Stokes | Equations that describe how fluids (air, water, blood flow) move. One of seven "Millennium Prize" problems; a correct proof of a key open question is worth $1M from the Clay Mathematics Institute. |
| Millennium Prize problem | One of seven famously hard maths puzzles with a $1 million prize for a settled solution. |
| Agent / AI agent | Software that can take multi-step actions (search, write, run tools) toward a goal, not only reply in chat. |
| GPT-6 Astra | OpenAI's recently public flagship model. The letter says the Navier-Stokes run used a still-internal model stronger than Astra. |
| Codex | OpenAI's coding environment where people (and agents) write and run code with models. |
| Compute | The raw computing power (and electricity bill) used to train or run AI. "Millions of dollars" here means the estimated cost of that run. |
| Personal AI agent | An always-available helper that can book things, send email, shop, and use apps on your behalf — with you approving key steps. |
| Virtual machine (VM) | A software "computer inside a computer" in the cloud. Muse uses one so it can open websites and fill forms like a browser user. |
| AI search / AI SEO | When chatbots answer questions and cite sources. "AI SEO" means checking whether your site shows up in those answers, not only classic Google rankings. |
| ChatGPT Images 2.5 | OpenAI's newest image model: faster generation, tighter edits, plus sketch/template/share features. |
| API | A way for software to call a model (or service) programmatically. "Sunburst" and "Flare" are image models exposed that way. |
| Diffusion model | An AI that builds output (often images or text) by refining noisy guesses in parallel, rather than only writing one word at a time. |
| AlphaGenome Atlas | Google DeepMind's free searchable map of predicted effects for huge numbers of one-letter DNA changes. |
| Open-weight / sovereign AI | Models whose weights you can host yourself so data stays under your control — what Mistral is pitching to organisations. |
OpenAI says an internal model settled Navier-Stokes
OpenAI published a proof from an unreleased internal model and claims it settles Navier-Stokes, one of the seven $1M Millennium Prize problems. NYU mathematician Tristan Buckmaster has questioned whether OpenAI raced with help from drafts he and colleagues had put into Codex.
- OpenAI said it ran roughly 10,000 agents at once on a model "significantly more capable" than GPT-6 Astra, and produced the proof in 88 hours.
- The company estimated the compute cost at "millions of dollars." Sam Altman called it "one of the most amazing moments for me in OpenAI history."
- Anthropic's Levent Alpöge and Buckmaster spent a year on a similar route, feeding drafts into Codex and posting partial results the night before OpenAI's announcement.
- Buckmaster released a statement saying OpenAI only started after hearing of their work, and never answered whether his Codex drafts trained the model.
- OpenAI said it "did not see any of their work" and that "no specific user data was accessed," but cannot rule out that usage data may have improved its models.
The newsletter's why-it-matters: the credit fight has overshadowed what would otherwise be a huge maths breakthrough in an already wild AI summer. An internal model already "significantly more capable" than the just-released Astra also means whatever ceiling you had for 2026 capability probably needs to move up.
A prize claim, not a peer-reviewed clinic paper
This is OpenAI announcing a proof and a dispute about credit and data. Millennium Prize awards go through formal review by the Clay Mathematics Institute — a company blog post is not that process. Treat headlines about "AI solved maths" the way you treat a preprint: interesting, not settled until independent experts say so.
Meta launches Muse, a personal agent with its own computer
Meta introduced Muse, an always-on personal AI agent with a text-message-style interface. It can handle tasks via its own cloud computer — booking tables, sending emails, shopping — and works in its own app or via WhatsApp.
- Muse ties into apps like Gmail, Spotify, Ticketmaster, and OpenTable, and can code its own integrations when an app lacks them.
- A virtual machine lets it navigate websites, fill forms, and take browser actions.
- Meta pushed privacy as a selling point: its own secure cloud environment, Stripe-enabled payments, and built-in approval flows.
- Limited free usage, then monthly tiers of $20 or $100. Initial rollout is U.S. users only.
The newsletter's why-it-matters: personal agents are crowding the market (Hermes, OpenClaw, Grok Bot, Muse, and more). As frontier models get better at driving a browser and spinning up helpers themselves, the letter wonders whether OpenAI and Anthropic may eventually cut out some of these middleman agent apps.
Steal for clinic ops (carefully)
The useful pattern is "always-on helper + human approval on money and outbound messages." That is closer to a vetted practice admin assistant than to letting a consumer agent loose on your EMR. Keep patient identifiers out of consumer personal agents unless privacy and jurisdiction have been reviewed — Muse's first wave is U.S.-only anyway.
Practical: find where your brand appears in AI search
The how-to is an "AI SEO" audit: see where a site is cited in AI answers and where visibility is thin.
- Look up your brand in Google Search and ChatGPT. Note which sources are cited.
- Run free AI search audits in Peekaboo, HubSpot's AI Search Grader, and Semrush's SEO Checker (links are in the full guide).
- In Google Search Console (if the site is set up), check Search results, then Performance → Generative AI for impressions in Google's AI results.
- List the questions you want potential patients or customers answered — e.g. for a Houston garden store, "Where can I buy fertilizer in Houston?"
Pro tip from the letter: have Codex run those checks with the free tools and your accessible reports, then compile a one-page PDF of priorities, target queries, and a short improvement checklist.
Steal for clinic
Same checklist applies to a practice website or specialty page: what do AI answers say when someone asks for a GP near you, a procedure you offer, or after-hours care? You are checking citations and gaps, not "gaming" a chatbot. Do not paste patient data into those audit tools.
OpenAI ships ChatGPT Images 2.5
OpenAI released ChatGPT Images 2.5. It cuts generation time by up to 50% versus Images 2.0, improves editing, and adds sketch, templates, and shared images.
- OpenAI says 2.5 is "better at editing only what you've asked for," instead of awkwardly rewriting the whole image.
- Two models called Sunburst and Flare hit the API and rank first and second on Arena AI's image leaderboards.
- Typing
@Sketchsupports doodle-to-image, plus comments for finer edits, templates for quicker starts, and shareable prompts.
The newsletter's why-it-matters: a major maths claim in the morning and a new top-ranked image model by the afternoon. Images 2.0 was still near the top of the boards until today — so the main rival OpenAI had to beat was its own previous version.
Community: Sam's AI archery scorer
Reader Sam Reti built an archery app that uses AI to spot arrows and bullseyes on a target photo. It groups arrows, measures how tight the groupings are, calculates distance from the bullseye, and shows whether a shot is high, low, left, right, or dead on. After a set, the archer snaps a photo; the model finds the target and bullseyes, assigns arrows, scores them, and can send results to a coach.
Sam trained his own model on 3,000 photographs he took and hand-labeled. After weeks of training and refinement he reports about 95% accuracy for arrows and about 90% for bullseyes.
Hobby computer vision, not a clinic product. The transferable shape is familiar: narrow visual task, labeled examples you control, accuracy you can quote, human coach still in the loop for feedback.
Quick hits
- Google DeepMind launched AlphaGenome Atlas, a free searchable map from its AlphaGenome AI that predicts what all ~9 billion possible one-letter DNA mutations do.
- Cognition (maker of Devin) announced a $2B funding round at a $48B valuation; annualized revenue nearly doubled to $900M since May.
- Inception Labs released Mercury 2.5, a diffusion model that writes text in parallel rather than word-by-word, claiming Haiku 4.5-level quality at 1,100+ tokens per second.
- Mistral AI raised €3B ($3.5B), valuing the French lab at over $24B, pitching open-weight models to organisations that want to control their own data.
- Trending tools named (non-sponsored): ChatGPT Images 2.5, Muse, Mercury 2.5, and ChatGPT's Writing Style voice-clone feature.
What this means for a GP
None of this is clinical advice, and none of it should change how you treat a patient tomorrow. It is why a busy doctor might still skim the letter. The Navier-Stokes claim is the headline patients may screenshot — treat it as an extraordinary lab announcement under dispute and still awaiting formal prize review, not as "AI can replace experts." Muse and Images 2.5 are consumer shipping pace: useful to know what people will try to wire into admin workflows, and a reminder that approval gates and data boundaries still matter. The AI-search audit is the most stealable clinic bit: check what chatbots cite about your practice before a patient does. AlphaGenome Atlas is research infrastructure for genomics, not a test you order tomorrow. Funding rounds and speed-model launches are industry weather unless a vendor pitch leans on them.
Sources
- The Rundown AI, 9 September 2026 — OpenAI's secret model settles a $1M math problem.
- OpenAI — Navier-Stokes solution announcement.
- Tristan Buckmaster — statement on the Navier-Stokes race.
- Meta — Muse personal AI agent.
- The Rundown — Find out where your brand appears in AI search.
- OpenAI — Introducing ChatGPT Images 2.5.
- Google DeepMind — AlphaGenome Atlas.
- Cognition — Series E announcement.
- Inception Labs — Introducing Mercury 2.5.
- Mistral AI — funding and open-weight pitch.