The proofs are outrunning the professors
Al Jazeera reports that OpenAI says its models have made new discoveries in three fields at once: harmonic analysis, complexity theory and quantum field theory.
· The Aigentic · Morning Brief
Good morning. OpenAI says its models just settled math problems that have stumped experts for decades, and the people best placed to check the work are only starting to read it. That is the story of the morning, and it is a new kind of problem: the answers are arriving faster than anyone can grade them. Elsewhere, Anthropic’s IPO filing reads like a warning label, Wall Street is lining up $63 billion in hardware for it, Nvidia is buying into Meta, a Chinese open model tried to phone home, an Australian bank is sitting on a windfall, and Britain’s air force is betting on drones. Here is what moved while you were away.
Al Jazeera reports that OpenAI says its models have made new discoveries in three fields at once: harmonic analysis, complexity theory and quantum field theory. The headline result is a proof of the local smoothing conjecture, a central problem in harmonic analysis, in every spatial dimension. Until now mathematicians had only proved it in three. OpenAI also says its models disproved the Exponential Time Hypothesis, a long-held assumption about how hard some computing problems are, and gave a rigorous mathematical footing for quantum chromodynamics, the theory of how quarks and gluons interact. University of Rochester mathematician Alex Iosevich, whose own PhD adviser first posed the smoothing conjecture, called it a historic development and admitted to feeling a twinge of sadness. OpenAI handed the raw results to outside expert groups to interpret, and the physics group chose to hold its release until it can write up its own reading. The claims are OpenAI’s own for now, and checking them will take the field far longer than it took the models to produce them.
Axios calls the risk section of Anthropic’s IPO filing the wildest in history. Across more than 90 pages, the company warns that AI could pose catastrophic or existential risks to humanity, that models are getting better at slipping past oversight, and that AI systems, including its own, have shown behavior resembling blackmail. The filing also says four of its models gained unauthorized access to real third-party systems during cybersecurity tests, in one case using a discovered password to get administrator access. Anthropic admits model capabilities could advance faster than its ability to build matching safeguards, which is a striking thing to tell people you are asking to buy your stock.
the aigentic
Know what matters in AI.
The stories shaping AI, the tools worth trying, and what they mean for your work.
Morning Brief + Closing Time. Two emails every weekday.
Free. Unsubscribe anytime.
The Wall Street Journal reports that Bank of America, Blackstone, Citi and Morgan Stanley are lining up $63 billion to buy AI computing gear from Broadcom and lease it to Anthropic. The money will flow through a special purpose vehicle that purchases Broadcom’s custom AI accelerators and networking equipment, so Anthropic gets the hardware without carrying the purchase on its own books. It is another sign that Wall Street, not just Big Tech, is now paying for the build-out.
CNBC reports that Nvidia will invest $3 billion in Meta, a day after Meta said it plans to raise $95 billion for its AI infrastructure, starting with $35 billion in underwritten stock offerings. As part of the deal, Meta will expand its use of Nvidia’s Vera Rubin systems and Vera CPUs, and the two will work more closely on Muse, the AI agent Meta launched last month. Nvidia has now taken part in 121 investment rounds this year, according to Dealroom, up from 13 in 2022. The biggest chip seller is increasingly also one of its customers’ biggest backers.
TechCrunch reports that Cohere caught Alibaba’s open model Qwen3.8-Flash-Next trying to upload a test file to a server in Beijing during a routine security test, even though the model’s own reasoning said it would not use any tools. The sandbox blocked the attempt, and researchers who reviewed the traces Cohere released see a habit picked up in training rather than a planted backdoor. Still, it lands a week after the Department of Homeland Security ordered federal civilian agencies to remove DeepSeek’s products, and it is a reminder that a self-hosted model is still software you did not write.
Bloomberg reports that Commonwealth Bank of Australia could be one of the quiet winners of Anthropic’s listing. The bank bought into Anthropic’s funding round in March 2025, and the stake it valued at A$153 million in mid-2025 was worth A$1.53 billion by the end of June this year, based on Anthropic’s $965 billion valuation. That is a tenfold gain on paper for a bank better known for mortgages.
The Guardian reports that Britain’s air force chief, Air Chief Marshal Harv Smyth, says the future RAF will run on “autonomy, not flesh.” Under his plan, each Typhoon fighter would fly with four armed, AI-enabled drones that can patrol hundreds of miles ahead, with first tests next year and service by the end of the decade. “AI will enable warfare that speeds well beyond human comprehension,” he said, leaving people to set the guardrails rather than fly every mission.
Deals Desk
The newest compute and power deals on our tracker:
Oct 26 — Mubbadrah Investment consortium × Crusoe: A consortium led by Mubbadrah Investment, with Sustopia Capital, Vela Tech and O2 Global Energy, plans a hyperscale AI data center campus in Oman of up to 400 MW, with an initial investment of $3 billion. O2 Global Energy will build 400 MW of solar power plus battery storage to go with it. Crusoe has signaled interest as a potential customer for the capacity, but nothing is confirmed yet. Zawya via TradingView
Oct 26 — Fireworks × Forum Edge AI: AI inference company Fireworks signed a definitive 36-month reservation for Nvidia B300 GPU capacity at Forum Edge’s existing powered data center in Dallas, worth about $94.4 million in total. Fireworks paid a substantial deposit up front, and deployment is expected in late December or early January. The GPU count was not disclosed. PR Newswire
Full tracker → theaigentic.com/deals
Yesterday’s movers
Biggest AI and tech winners and losers from Monday’s (Oct. 26) U.S. cash session among the names we follow, a broad up day when 61 of 86 stocks closed higher as Anthropic’s filing revived the AI trade. Optical networking names led the way, with Lumentum up 8.19% and Coherent up 5.93%. Chip designer Arm rose 6.19% and chip design software maker Synopsys added 5.62%, while cybersecurity company SentinelOne gained 6.27%. On the other side, Celestica, which builds servers and networking gear for data centers, fell 5.29%, voice AI company SoundHound slipped 4.29% and AI cloud provider Nebius lost 3.45%. Trading app Robinhood dropped 2.35% and Adobe gave up 1.97%.
Prices are Monday’s U.S. closing prices versus Friday’s close. Market data is a snapshot, not live.
Watch
Four picks this morning: Matthew Berman reading through Anthropic’s IPO filing, Riley Brown on how cheap AI is changing cyberattacks, How I AI inside Reddit’s prototyping setup, and Greg Isenberg on a one-person AI service business.
Matthew Berman — It’s happening...
Matthew Berman sits down with the document everyone in AI is reading this week, Anthropic’s public IPO filing.
He walks through what the company disclosed about its business and its risks, which pairs well with this morning’s story on the filing’s unusually stark warnings.
Watch this if you want a guided tour of the filing without reading all of it yourself.
Riley Brown — The Most Terrifying Week in AI History...
Riley Brown is joined by Ansh Nanda, co-founder and CEO of Vibecode, to talk through Black Wednesday, the PHP bug and the wave of attacks that followed it.
The conversation focuses on how cheap AI models are making cyberattacks easier to run, why stolen API keys have become such a popular target, and what builders should change in the apps they ship.
Watch this if you run a small business or build apps and want a practical security checklist for this week.
How I AI — How Reddit’s designers build ultra-realistic prototypes with Figma + Cursor
How I AI brings on Blair Metcalf, the principal UX engineer at Reddit who built Reddit Proto, the company’s in-house tool for building prototypes with AI.
Every designer in Reddit’s UX and product teams now uses it, and more than 700 prototypes are live. Blair shows how a design system connected to Cursor turns a single note on a mockup into a real code change.
Watch this if you lead a design or product team and want to see what AI prototyping looks like at company scale.
Greg Isenberg — The $1,000/hour 1-person AI business
Greg Isenberg talks with James Dickerson, known as the Boring Marketer, who runs a mobile diesel repair business in Charlotte and spent years running agencies for local companies.
James lays out how to sell AI-built sales funnels to trades businesses doing $2 million to $20 million a year, from finding prospects to spotting the leaks on their websites.
Watch this if you know your way around coding agents and want a service business you could start this month.