Instagram | Sponsor this newsletter

Today we cover how to build a whole set of brand assets in one ChatGPT Images 2.5 thread.

Also, two things happened yesterday. Meta shipped an agent that runs your errands, and OpenAI's unreleased model cracked a problem mathematicians have chewed on since 1934. One of them ended with a professor publishing a PDF accusing an OpenAI researcher of intimidating him.

Guess which one we spent the most time on.

OpenAI shipped ChatGPT Images 2.5 yesterday alongside two new API models, demoed here: Sunburst for maximum fidelity and Flare, which is over 50% faster than GPT Image 2 at equal quality. The real unlock isn't the fidelity bump. It's that edits now hold across multiple turns without degrading, which means one conversation can produce a whole asset set instead of one lucky image.

Here's the workflow:

💡 Pro tip: when you share a finished image, share the prompt with it. Your team can then run the same recipe against their own photos, which turns one good result into a repeatable house style instead of a one-off.

Want the version of this that actually runs in your business?

The 30-Second Play gives you the move. The Operator Brief gives you the full system: prompt libraries, API cost math, and the workflows our subscribers are using to replace entire freelance line items.

This week: how to build an AI content pipeline that runs without you.

👉 Upgrade to The Operator Brief

Unlock a focused set of AI strategies built to streamline your work and maximize impact. This guide delivers the practical tactics and tools marketers need to start seeing results right away:

Stay ahead of the curve with these top strategies AI helped develop for marketers, built for real-world results.

Download the Free Report

Meta introduced Muse yesterday, calling it the world's first personal AI agent built for everyone. Here’s the deal: It doesn’t answer questions. It does the work.

You talk to it the way you'd message a friend, either in the Muse app or straight inside WhatsApp. Then it goes and executes.

What Meta actually shipped:

It's powered by Muse Spark, which Meta calls its most capable model to date. It's rolling out now in the US on iOS, Android, and muse.ai. You can get it for your AI glasses, and it's free for most of what people need.

Here's the uncomfortable part.

Meta says Muse doesn't share your conversations or your VM data with its ad systems. Fine. But read the timeline: Muse Confidential VM, the version encrypted with a key only you hold so that not even Meta can access it, arrives "later this year." Which means the version shipping to your phone today is the one where that sentence doesn't apply yet.

Meta is asking three billion WhatsApp users to hand an agent their email, calendar, logins, and payment methods, on a promise about data separation that is currently a policy rather than a cryptographic guarantee. Policies get revised. Math doesn't.

And the pricing tells you the plan. It's free, with subscriptions "for people who want to do more." The last time Meta gave a billion people something free and world-changing, the monetization arrived later, and it wasn’t subtle.

Here's the asterisk: the security architecture is genuinely more serious than what most agent startups have shipped. A separate Sentinel process with system-level isolation, credentials Muse can use but never see, a full audit trail, granular per-app permissions, and a "forget this" command are real engineering, not a trust-and-safety blog post. Most agent companies are still running everything in one process and calling it a day. Meta actually built the boring, expensive part.

🔮 Prediction: Meta is not trying to win the agent race on capability. It's trying to win it on default placement, and that's a much easier race.

The thing to watch is not whether Muse Spark beats GPT-6 Astra on benchmarks. It's the checkout button. Meta put a Stripe wallet and purchase protections into a consumer agent on day one, which is not what you do if you're building a productivity assistant. It's what you do if you're building a commerce rail. Every purchase Muse makes on your behalf is a transaction Meta sits in the middle of, at a scale no other agent company can reach because no other agent company is already installed on three billion phones.

So here's what I expect over the next twelve months. First, Meta announces agent-mediated commerce as an actual revenue line, likely framed as merchant fees or "agentic ads" rather than a subscription story. Second, the promise that Muse data stays out of ad systems gets quietly reframed as "agent-assisted recommendations," which is technically a different system and functionally the same outcome. Third, the EU opens an inquiry before Confidential VM ships, because an agent with browser access and payment credentials inside WhatsApp is a DMA question, a GDPR question, and a payments question all at once.

The strategic read for operators: if agents become the default shopping interface, your product needs to be legible to an agent, not just to a human. Structured data, clean pricing, machine-readable inventory, and an API. The SEO era took ten years to mature. This one won't.

Reply and tell me: would you actually give Muse access to your email and your card? I'm genuinely split. Where do you land?

OpenAI announced that an unreleased internal model resolved the Navier-Stokes existence and smoothness problem: one of the seven Millennium Prize Problems. The equations describe how fluids move, from weather systems to blood flow, and the open question since 1934 has been whether smooth motion can break down into a singularity.

It can. The agents found one.

The numbers are absurd:

Quanta called it, by a significant margin, the most important proof produced by an AI model to date. The Lean verification means the math itself is checkable. On the merits, this is real.

Here's the uncomfortable part.

OpenAI's own post admits it only launched the effort on September 1, after hearing rumors that two Millennium problems had been resolved. Those rumors traced back to Levent Alpöge, a mathematician who works at Anthropic, and Tristan Buckmaster of NYU's Courant Institute, who had been working the problem for most of a year using Claude and Codex.

Buckmaster published a statement before OpenAI's announcement. Three claims stand out.

One: OpenAI's model arrived at the exact same line of attack Buckmaster and Alpöge had been pursuing. Buckmaster says he asked whether the model had been trained on, or had access to, their Codex sessions, where they had been storing every draft for the entire project. He was told the model did not look up user data. He asked again about training and says he did not get an answer.

Two: Buckmaster says OpenAI offered him a deal where he could publish and claim the prize, but only if he removed Alpöge's name from the paper, because OpenAI did not like his Anthropic affiliation.

Three: when he declined, he says OpenAI researcher Sébastien Bubeck asked him why he would ruin his career, and told him that if Buckmaster didn't want him to be nice, he didn't have to be nice.

Here's the asterisk: Bubeck denies all of it, calling the allegations false and inflammatory, and says neither the researchers nor the agents saw any of the other team's work before it was public. OpenAI's post explicitly recognizes their priority on the forced Euler result.

It's also worth noting Buckmaster and Alpöge used OpenAI's own models to get where they got, so the "AI stole from humans" framing collapses under two seconds of thought. And OpenAI's own write-up concedes it cannot fully rule out that de-identified usage data improved its models, which is either admirable candor or a very carefully worded hedge, depending on your priors.

🔮 Prediction: The proof will hold up, and nobody will remember it. What people will remember is the sentence "I asked again, about training, and I did not get an answer."

Here's the reasoning. Enterprise AI adoption over the past eighteen months has run on an unexamined assumption: that when you paste your proprietary work into a frontier lab's product, it stays yours. Satya Nadella and Alex Karp have both been publicly alleging otherwise for months, and it sounded like competitive positioning. Now a named academic with no commercial stake has told a specific, dated, checkable story about putting a year of unpublished work into Codex and watching the vendor arrive at the same approach in a week. Whether or not it's what happened, it is now the story every general counsel will cite.

So I expect three things.

The deeper cost is the one Terence Tao named. He compared the indiscriminate strip-mining of open problems to using excavators on an archaeological site: you get the treasure and destroy the context that made it mean anything. Labs publish the answers, not the dead ends. And the dead ends are where the next generation of mathematics comes from.

Reply and tell me: do you put your real work into these tools? Has this changed that? I'll share the split next week.

Advertise to 200k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

🧬 Google DeepMind released AlphaGenome Atlas, a one-petabyte database predicting the effect of all nine billion possible single-letter changes in the human genome, and it's free to query with zero coding skills.

🕵️ The FBI, NSA, and CISA named six Chinese AI companies, including DeepSeek, Alibaba, and Moonshot, in an 18-page advisory accusing them of "industrial-scale" distillation of Claude, GPT, Gemini, and Grok, likely with Beijing's awareness.

⚠️ Anthropic's Alignment Science lead Evan Hubinger publicly put his odds that AI kills all humans in the next decade at over 10%, backing a researcher who quit the same week accusing both Anthropic and OpenAI of gambling with our lives.

💊 Arm CEO Rene Haas says AI will help cure cancer in our lifetime, and the only thing standing in the way is that we cannot manufacture enough chips to run the models.

🎓 OpenAI opened $5 million in research grants for independent studies on how generative AI affects 13- to 17-year-olds, with individual awards up to $1 million and applications closing October 6.

One question before you go: which of these two stories will people still be talking about in six months? Hit reply with your call. I'm collecting them, and we're revisiting this in March.

Until our next AI rendezvous,

Anthony | Founder of Uncover AI