AI in 15 — July 15, 2026

The man who runs Google DeepMind — one of maybe four people on Earth closest to building artificial general intelligence — just published a manifesto asking the United States to build a watchdog that could order his own industry to slow down. Before the end of this year.

Welcome to AI in 15 for Wednesday, July fifteenth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. And today has an unusual shape, Kate — two of the biggest names in AI, independently, on the same day, arguing the smart move is to put a governor on the machine.

It really is, Marcus. Our lead: Demis Hassabis wants a US-led global AI checkpoint operational by year-end. Then a run worth your time.

The people who scared everyone with "AI 2027" now have a hundred-page plan to push superintelligence out to 2040 — on purpose.

Oracle has lost half a trillion dollars, and the central bank of central banks is nervous.

The father of reinforcement learning bets his new company that today's chatbots are a dead end.

Two AI security stories that pull in opposite directions.

And Meta shops for chips that don't say Nvidia.

Lead story, Marcus. Hassabis published this yesterday. What is he actually asking for?

So this is a personal manifesto, Kate — titled "A Framework for Frontier AI and the Dawning of a New Age" — and Axios broke the details. Hassabis wants the United States to stand up a brand-new oversight body with real teeth: the power to screen the world's most advanced models before release, and — this is the sharp part — to coordinate an industry-wide slowdown if the danger signals start piling up. He calls it a security checkpoint for frontier AI. Industry-funded, staffed with world-class technical experts, answerable to the US government, with a majority-independent board he wants stacked with Turing Award winners.

And he wants this running by December?

Before the end of 2026, yes. And the scope is deliberately huge, Kate — it would apply to every frontier-class model, quote, "no matter their country of origin or whether they are open or closed." He also says, flatly, that AGI — a system with the full cognitive range of a human brain — is probably only a few short years away, and that we're standing in what he calls the foothills of the singularity.

Okay, so a sitting CEO is telling us AGI is nearly here and please regulate me. Is that as noble as it sounds, or is there an angle?

That's exactly the tension worth sitting in, Kate. On the generous read, this is genuine risk management from someone who understands the technology as well as anyone alive. On the skeptical read — and this was well represented over on Hacker News — a US-led regime mostly regulates US labs, while China keeps sprinting. So you set a good standard that doesn't actually bind the competitor you're most worried about.

And the AGI-is-close claim?

Probe it, Kate. Notice that his commercial incentive and his safety argument point the exact same direction. "AGI is imminent and dangerous" makes DeepMind's models sound world-changing and makes a screening body sound urgent — a body that a well-resourced incumbent is very well placed to comply with, and a smaller rival maybe isn't. It can be sincere and a regulatory moat at the same time. Both things fit in the sentence.

Story two, Marcus, and it rhymes with the lead — but it's a totally separate group of people. The AI 2027 forecasters are back.

Right, and let's not force these two together, Kate — they landed the same week by coincidence. This is the AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo. Last year their viral "AI 2027" scenario convinced a lot of people superintelligence was arriving fast and messy. Their follow-up is roughly a hundred pages, called "AI 2040: Plan A," and the striking thing is the reversal. The same team that war-gamed a runaway 2027 now argues the wise move is to deliberately push superintelligence out to around 2040 instead of racing there by 2030.

So they talked themselves down off the ledge.

Sort of — but they're careful, Kate. They say explicitly this is a recommendation, not a prediction. What they think should happen, not what will. And the spine of the plan is a deal: by 2029, the US and China agree to make frontier AI research transparent to each other and to independent monitors, backed by verification and a concept they call — I love this — "mutually assured compute destruction." The credible ability to disable your rival's data centers if they cheat. A Cold War deterrent for the AI age.

That is an enormous ask. Do they think it happens?

They put the odds of Plan A actually being adopted at three to fifteen percent, Kate. So even the authors don't bet on it. And the obvious hole — which they don't hide — is that it requires Washington and Beijing to open their most strategic labs to each other and treat the threat of blowing up each other's data centers as stabilizing rather than terrifying. The value here isn't feasibility. It's that "slow down" is finally a detailed engineering-and-treaty proposal instead of a bumper sticker. Policymakers now have something concrete to argue with.

Story three, Marcus, and this is the one hitting people in the portfolio. Oracle is in freefall.

Nearly fifty percent down since its June first high near two hundred fifty dollars, Kate — that's roughly half a trillion dollars in market value gone. It knocked co-founder Larry Ellison from the world's number two fortune down to number eight, behind Nvidia's Jensen Huang. His net worth dropped about a hundred and twenty-five billion in six weeks.

What spooked everyone? Oracle was the AI darling a month ago.

The financing, Kate. Oracle plans to spend up to ninety-five billion dollars building AI data centers for customers like OpenAI. It ended its fiscal year with a staggering six hundred thirty-eight billion in remaining performance obligations — booked future work — up three hundred sixty-three percent in a year. And roughly three hundred billion of that is tied to OpenAI alone. That's enormous concentration risk: your backlog leans on one customer that has its own giant capital commitments to meet. S&P actually cut Oracle's credit rating over it.

And this isn't just Oracle, is it.

No, and that's why it's a story and not a footnote, Kate. Newly public SpaceX has slid to a record low near a hundred thirty-nine dollars — down about thirty-eight percent from its peak, below its IPO price — as the orbital-data-center AI narrative wobbles. And the Bank for International Settlements — the central bank for central banks — put out a bulletin this week warning that AI infrastructure is increasingly funded by debt rather than by cash flow. They flagged AI financing as a potential systemic risk.

So is this "the technology is failing"?

No — and that's the important distinction, Kate. Nobody's saying the models got worse. This is the market re-rating the build-out's debt structure in real time. For two years the only AI market story was up and to the right. Now the boring question — who actually pays for the data centers, and what happens if a marquee customer can't meet its obligations — is finally getting priced. Faster is no longer the only story investors believe.

Story four, Marcus, and it's a genuine dissent from inside the field. Richard Sutton started a new company to bet against the whole current playbook.

And when Sutton says it, you listen, Kate — he's the Turing Award-winning pioneer of reinforcement learning, one of the founding figures of the field. He's left John Carmack's outfit to found Oak Lab, and the thesis is blunt: today's large language models are a dead end. He's building something called the OaK architecture — Options and Knowledge — agents that learn continuously from real-time experience, on the job, without storing or replaying giant datasets, on far less compute.

Unpack the disagreement for me. What's he actually rejecting?

The entire pretrain-on-the-internet paradigm, Kate. His argument rests on what he calls the Big World Hypothesis — the real world is just too vast and complex to compress into a static training set. So genuine intelligence has to be built at runtime, learning by doing, not pre-loaded by reading everything humans have ever written. His wager, put simply: intelligence comes from doing, not from reading. And that's a direct shot at what OpenAI, Anthropic, and Google DeepMind are all scaling.

Is he a lone voice, or is there a signal here?

There's a rhyme, Kate. A London robotics startup, Humanoid, showed this month that trial-and-error reinforcement learning — not more pretraining — is what pushed real-world reliability up on factory tasks. Bin-picking success from eighty percent to ninety-eight, on their own numbers, so treat those as vendor figures. But the pattern is the same idea Sutton's staking a company on. When the most decorated name in RL says the LLM era is a detour, that's not a critic on the sidelines. That's a real fork in the road.

Two security stories now, Marcus, and I love that they point opposite ways. Start with Cursor.

So security firm Mindgard went full-disclosure on a Cursor vulnerability after, they say, more than seven months of trying to report it quietly, Kate. The flaw is ugly in its simplicity: on Windows, when Cursor opens a project it goes looking for git binaries, including in the workspace itself. So if a repo contains a malicious file named git-dot-exe in its root, Cursor just runs it — no click, no prompt, no warning. Arbitrary code execution the moment you open a poisoned repo.

Clone a repo, get owned.

Exactly, Kate. And Mindgard says it first reported this back in December, the vendor initially closed it as "Informative," and the issue survived through a hundred ninety-seven-plus releases since. When Dark Reading reached Cursor on Monday, they said they're addressing it — but gave no fix timeline. There's a fair caveat: some argue the current directory shouldn't be on the search path in the first place, so it's partly a Windows behavior, and an attacker still has to get you to open their repo. But the sharper story is the disclosure process — a vendor moving fast on features and slow on a reported remote-code-execution bug.

And the flip side — OpenAI's Codex.

This one's about transparency going the other way, Kate. OpenAI changed its Codex CLI so that when a parent agent delegates to sub-agents using its Sol or Terra models, the prompts between them are now encrypted. On your machine, Codex stores only encrypted content — the readable field is left empty — and it's decrypted server-side. So you can no longer see what tasks your own agents are handing each other. Developers who built local tools to audit their coding sessions found them suddenly broken.

Why would OpenAI do that?

The charitable read is stopping unauthorized API pooling and reverse-engineering of their multi-agent orchestration, Kate. The worried read — and it's legitimate — is that there's now an encrypted set of instructions running on your machine that you're not allowed to read, which is a real problem if you need transparency for compliance, debugging, or safety review. And notice the irony against today's lead stories: while Hassabis and the AI 2040 crowd argue for more transparency in frontier AI, a leading lab just made its tooling more opaque. Small change, big principle — as agents act on your machine and delegate to each other, who gets to see the instructions?

Last hit, Marcus. Meta is talking to Google about chips.

And this pokes right at Nvidia's soft spot, Kate. Meta is reportedly in talks to deploy Google's custom TPU chips in its own data centers starting in 2027, and possibly renting them via Google Cloud as early as next year — a multiyear, multibillion-dollar arrangement. Meta is one of Nvidia's biggest customers, with up to seventy-two billion dollars in planned AI spending. So any real shift toward Google silicon is a notable dent. Google Cloud people reportedly think broader TPU adoption could capture up to ten percent of Nvidia's annual revenue.

So is Nvidia's grip slipping?

Not slipping — being tested, Kate. Nvidia's dominance rests partly on the assumption the hyperscalers can't or won't switch. And to be fair, Google and Nvidia both went out of their way this week to emphasize their multi-decade partnership — Google noted GPUs are still the most popular chips on its cloud. So this isn't Meta abandoning Nvidia. It's the biggest buyers building a second source and leverage. And tie it lightly to the Oracle story: if the whole infrastructure trade is getting re-rated, then who controls the chips — and at what margin — is exactly what's under the microscope.

One to watch tomorrow, Marcus.

The Hassabis watchdog, Kate — specifically whether any other frontier lab, OpenAI or Anthropic, publicly backs it, hedges, or rejects it. That reaction will tell you whether this is a real governance moment or one CEO's op-ed.

Agree, or counter?

Small counter, Kate. Watch Oracle and the AI-infrastructure names instead. The market, not the manifestos, may be the thing that actually sets the pace this year.

That's your AI in 15 for today. See you tomorrow.