Early Access

00:00:04 Here's how Tuesday went. President Trump signed an executive order on artificial intelligence — and the most telling detail is how he signed it. The plan, for weeks, was a big room. Silicon Valley's top chief executives standing behind him, cameras rolling, the president putting his name to a framework for overseeing the most powerful AI models in the country.

00:00:25 That's not what happened. He signed it privately. No CEOs and no staging. And the order he signed was a smaller thing than the one he'd planned to sign back in late May. Let me walk you through what it actually does, because the gap between the May draft and the June signature explains everything else.

00:00:43 The order asks certain AI companies to voluntarily submit their newest models to the federal government for testing thirty days before they ship to the public. Thirty days. The earlier draft had said up to ninety. Industry had pushed back hard — some of them wanted something closer to two weeks — and the number moved their way.

00:01:02 Rebecca Bellan at TechCrunch reported that Trump had been slated to sign the more demanding version in late May, then delayed it after pushback that included David Sacks, the venture capitalist who served as his White House AI czar. The president's own stated reason for not pushing harder: he didn't want to do anything that would get in the way of American firms leading against China.

00:01:25 That sentence sets the ceiling here. Every time a rule gets close to binding, the China argument pulls it back open. Now, the word everyone's circling is voluntary. And the order goes out of its way to make voluntary mean voluntary. I'm going to read you the line directly, because the lawyers clearly fought over it.

00:01:43 Quote: 'Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.' End quote.

00:01:59 Read that twice. The order doesn't just decline to require a license. It pre-emptively forbids anyone from reading it as a license. That's the carve-out the industry wanted, written into the text in ink. The order does two other things. It tells the Department of Justice to treat AI-assisted crime — hacking, unauthorized access — as a high-priority enforcement area.

00:02:21 And it directs the cybersecurity agencies to harden the federal government's own communications and staff up a technical corps inside government. So it's not nothing. It moves real machinery. But the headline action — the government getting an early look at frontier models — rests entirely on companies choosing to show up.

00:02:40 Ryan Fedasiuk, who works on AI and national security policy, gave the fairest read I saw. He called it, quote, 'a common-sense, light-touch framework designed to get the federal government's gears in motion.' He also flagged what's missing: it's light on the details of what voluntary review even involves, and there's no safe harbor — no promise that a company handing its model over gets any legal protection in return.

00:03:05 So you have an ask with no teeth and no reward. A reply under his thread, from policy analyst Divyansh Kaushik, was blunter. The window is up to thirty days, he said — and he predicted it won't bind in practice any more than the ninety-day version would have. When the people who follow this closely are arguing about whether the soft number is soft or extremely soft, you already know roughly where the leverage sits.

00:03:30 So if there's no license and no safe harbor, what is the government actually trying to build here? This is where the order gets more interesting than the headline suggests, and where I'd slow down. Underneath the voluntary framing, there's a real piece of state machinery being stood up.

00:03:46 Within sixty days, the order directs the National Security Agency to build and maintain a classified benchmarking process. It's a way to measure the advanced cyber capabilities of AI models, and to set the threshold at which a model gets designated — in the order's language — a 'covered frontier model.' Sit with that for a second.

00:04:04 The thing that decides whether your model is powerful enough to fall under federal interest is a classified test. You won't see the benchmark. The companies won't fully see it either; the order says the assessments get shared with developers 'as appropriate.' So the definition of the most consequential category in the document — what counts as a frontier model worth the government's attention — lives behind a clearance.

00:04:28 Once a model is covered, the voluntary part kicks in. A developer can engage the government to confirm the designation, give it up to thirty days of pre-release access under confidentiality, and then this — 'collaborate with the Federal Government to select trusted partners that will have early access to covered frontier models to promote secure innovation and strengthen the cybersecurity of critical infrastructure.' Trusted partners.

00:04:53 The government doesn't just want to look at the model itself. It wants a hand in deciding who else gets an early look. And the order names who that's for. It talks about routing access to these models out to agencies, to state and local authorities, and to operators of critical infrastructure — and the examples it picks are revealing: rural hospitals, community banks, local utilities.

00:05:15 Not the Fortune 100. The places that can't afford a security team. The read I'd offer is that the government is trying to use frontier models as a defensive distribution channel — get the most capable cyber-defensive tools into the hands of the soft targets before someone uses an equally capable model to attack them.

00:05:32 That's a coherent goal. It's arguably a good one. But notice it depends, start to finish, on the labs agreeing to play. The benchmark is the government's. The models are the companies'. And the bridge between them is a voluntary handshake. That HackerNews thread on the order text had a comment I thought cut to it.

00:05:50 A reader named droidjj pruned the legalese down and asked the obvious question: 'So… it's entirely voluntary? This has no teeth, unless I'm reading it wrong.' He's not reading it wrong. But I'd push on the framing a little. 'No teeth' assumes the goal was to bite.

00:06:05 If the goal was instead to build the apparatus — the benchmark, the clearinghouse, the partner list, the relationships — without fanfare, so the binding version becomes easy to add later, then the soft order is doing exactly what a first step does. December's order, the one directing a single national rulebook to preempt state AI laws, already showed this administration prefers to centralize first and specify later.

00:06:28 Watch the sixty-day mark. The benchmark and the framework are due then. That's where you'll learn whether this is the groundwork for something real or a press release with a comment period.

00:06:39 Here's the thing that makes the executive order land strangely. The day before Trump signed it, Anthropic announced it was already running the arrangement the order describes — just without the federal government in charge of it. Anthropic said this week it's extending Project Glasswing, its program for finding and fixing critical software vulnerabilities with AI, to roughly 150 organizations across more than fifteen countries.

00:07:05 At the center of it is a model the company calls Claude Mythos, which it describes as its most powerful yet — able to identify thousands of zero-day vulnerabilities over a span of weeks. Back in early April, fifty initial partners, including the U.S. government, got access to a preview to scan their own codebases.

00:07:23 Now the circle widens to cover power, water, healthcare, communications, and hardware — the sectors that weren't well represented in the first round. Who's in the new group tells you what this actually is. According to the Financial Times, the organizations getting Mythos access include NATO, the European Union's cybersecurity agency ENISA, the identity-security firm Okta, and the South Korean giants Samsung, SK Hynix, and SK Telecom.

00:07:49 The countries lean exactly where you'd expect a Western security perimeter to lean: Australia, Canada, France, Germany, Italy, Japan, India, New Zealand, South Korea, and more. Anthropic's own framing of the stakes: quote, 'What each partner has in common is that a successful attack on their codebase could be catastrophic.

00:08:09 For most partners, we estimate that a major attack could affect more than 100 million people, with important ramifications for both global and national security.' End quote. So line the two stories up next to each other. The executive order, signed Tuesday, gestures at a future where the government helps select trusted partners to get early access to frontier models and harden critical infrastructure.

00:08:33 Anthropic, on Monday and Tuesday, simply did it — picked the partners, drew the perimeter, set the terms, and handed a frontier model's defensive power to NATO and an EU agency and a list of allied national champions. The government finally asked for a look at the most powerful models.

00:08:50 The labs had already been handing that look to NATO for two months. When I say the leverage doesn't sit in the order, this is what I mean. The capability and the relationships already live inside the companies. The state is trying to get a seat at a table the labs built and are still hosting.

00:09:07 Two more things round this out. First, the timing: this came a day after Anthropic confidentially filed for an initial public offering, following a sixty-five-billion-dollar round at a valuation near a trillion dollars. A company about to ask public markets for money is also positioning itself as indispensable to allied-government security.

00:09:27 Those two facts aren't unrelated, and the prospectus, when it surfaces, is where you'll see how much of Anthropic's story depends on being the West's security vendor. Second, this isn't a monopoly play yet — Anthropic itself says it expects rivals to field models as capable as Mythos soon, which is part of why it's racing to set the norms now.

00:09:48 OpenAI has already countered with its own cyber-focused model in partner testing. The contest to be the trusted vendor of national security is on, and it's running well ahead of the rule that's supposed to govern it.

00:10:01 Let me move to a different room, because two things happened in medicine on Tuesday that point at the same shift from different angles. The first: Microsoft and Mayo Clinic announced a partnership to build an AI model trained on Mayo's medical data, with plans for a consumer-facing health assistant and a set of tools for clinicians.

00:10:20 Clare Duffy reported it for CNN. I want to be careful here, because the detail is still thin and I haven't seen the governance terms — but the shape matters even before the fine print. Mayo Clinic is one of the most trusted names in American medicine, sitting on one of the richest clinical datasets in the country.

00:10:39 Microsoft is the platform partner threaded through the frontier-model economy. Put those two together and you get a model that carries Mayo's authority and runs on Microsoft's infrastructure, trained on patients' records, aimed at both patients and the doctors treating them.

00:10:56 The questions that raises are the ones I'd want answered before anyone calls this a breakthrough. What exactly is the model trained on, and did the patients whose data built it have any say? When the consumer assistant gives someone health guidance, who's liable if it's wrong — the hospital whose name is on it, or the platform that built it?

00:11:16 And what happens to every other hospital system that doesn't have a Microsoft partnership, when the trusted-brand medical AI belongs to the one that does? That last one is the consolidation question, and it's the same shape we just saw with Anthropic and national security.

00:11:32 The most capable tool ends up bundled with the most established institution, and the gap between the haves and have-nots in care delivery widens along the contour of who signed a deal with a frontier lab. The second door is Google DeepMind's. On Tuesday it unveiled Co-Scientist — a Gemini-based system that, in its own words, can 'generate, debate and evolve novel hypotheses for complex scientific problems.' The mechanism caught my eye.

00:11:58 It spins up a coalition of specialized agents and runs what DeepMind calls a 'tournament of ideas.' The system produces thousands of hypotheses, then holds structured debates to rank and refine them, checking each claim against the literature. The company says that over the past year it helped identify new drug targets for liver fibrosis and fresh approaches to ALS, the neurodegenerative disease.

00:12:22 It's now going out to individual researchers through a tool called Hypothesis Generation. What I appreciated was that the sharpest skepticism came from the replies, not from me. One researcher, Guilherme O'Tina, asked whether tournament-debate architectures push the system toward defensible hypotheses over truly novel ones — and noted those two things diverge more than people assume, which is exactly the distinction that matters for real discovery.

00:12:49 Another, posting as synabun, put it more dryly: a multi-agent debate to rank hypotheses is great, quote, 'until three of the agents confidently agree on the wrong one. Still better than peer review timelines.' That's the tension in one line. A machine that generates ten thousand plausible ideas is only useful if the ranking step can tell plausible from true — and we've spent the last two weeks on this show talking about fabricated citations and AI-generated science that looks right and isn't.

00:13:19 Speed at the front of the pipeline puts more weight, not less, on the verification at the back. I'd want to see Co-Scientist's liver-fibrosis and ALS leads survive a wet lab before I'd call them discoveries. But as a research instrument pointed at problems that have stalled for decades, it's the kind of thing I'm glad exists.

00:13:38 Pull the camera back, though, because every story so far has been about who's inside the room — which lab, which hospital, which allied government. Kenneth Rogoff wrote a piece on Tuesday about everyone who isn't. Rogoff is a Harvard economist and the former chief economist of the International Monetary Fund, so when he writes about distribution he's not speaking from the cheap seats.

00:14:00 His piece opens in San Francisco, where he describes a frenzy that, in his words, makes the Gold Rush 'look like a scavenger hunt' — top engineers offered compensation worth hundreds of millions to switch firms, young people who joined the right startup early contemplating retirement before thirty-five.

00:14:17 But underneath the gold, he finds anxiety: even the people inside fear their startup won't be the one that wins, and that they'll be automated out alongside everyone else. His real argument, though, is geographic, and it's where the piece gets sharp. The people displaced by AI, he writes, will overwhelmingly not live in the United States, and not in the handful of countries that secured a place in the AI supply chain — South Korea, Japan, Taiwan, the chip and memory economies.

00:14:44 They'll live in the places with no windfall to redistribute. Listen to how he frames the trap, because it's tighter than the usual hand-wringing. Quote: 'Countries that fail to carve out a place for themselves in the emerging AI economy risk ending up on the losing side of this century's most consequential economic transformation.

00:15:03 With no windfall profits to redistribute and no surge in tax revenues to finance universal basic income, they could find themselves with no way to cushion the shock of mass job displacement.' End quote. That's the mechanism that makes this worse than previous waves.

00:15:18 The standard comfort about automation is that the wealth it creates can be taxed and spread. Rogoff's point is that the wealth and the displacement land in different countries. India's enormous outsourcing industry — the call centers, the back-office processing, and the mid-level white-collar work — sits directly in the path of what these models automate first.

00:15:38 But the profits from automating it accrue to firms in California and the chipmakers in Asia. India gets the job losses without the tax base. He's even sharper on Africa and Latin America: how does an African firm compete in AI when hundreds of millions of people on the continent still don't have reliable electricity, which is the basic precondition for any of this?

00:15:59 How does a Latin American government finance data centers when savings rates are low and capital still remembers the last debt crisis? Some of these countries will benefit from selling the minerals the buildout needs — copper, lithium, cobalt, and the rare earths — but Rogoff notes resource wealth has a long history of being a curse as much as a blessing.

00:16:19 And here's where it connects back to yesterday. We spent time on Monday on Alphabet's eighty-billion-dollar share sale to fund its AI build-out — the largest equity raise of its kind, bidding to pull in more than the three biggest IPOs in history combined. The same rolling coverage that carried that number also carried a warning that AI threatens to drive up youth unemployment.

00:16:40 Those are the two ends of Rogoff's argument in a single news feed: record capital flowing to the firms inside the supply chain, and the first cracks in the labor market for everyone trying to enter it. The capability isn't the variable. The distribution is. And right now the distribution is running in exactly the direction that concentrates it.

00:16:59 Which brings me to Europe, where two small Tuesday stories say something about a continent trying to buy its way back into a game it's losing. The first: Bloomberg reported that the European Union's twenty-billion-euro plan to build five big AI data centers — the so-called gigafactories meant to give Europe its own frontier-scale compute — is floundering.

00:17:19 Delays, funding gaps, and partners getting alienated by the stumbles. I'd flag the contrast Bloomberg drew, because it's the whole problem in miniature: the same week the EU's public plan is stalling, SoftBank announced a seventy-five-billion-euro AI data center plan in France.

00:17:35 One private Japanese conglomerate is putting nearly four times the money into European soil that the entire European Union mustered for its flagship program — and moving faster. If you want to understand why Rogoff worries about who's inside the supply chain, watch the richest bloc outside the US struggle to stand up five buildings while the capital that actually moves comes from Tokyo and points at French real estate.

00:17:59 The second story is smaller and a little poignant. According to an internal email Politico got hold of, the European Parliament plans to replace Google with the French search engine Qwant as the default search tool on its computers. The framing is digital sovereignty — Europe wanting to run on European software inside its own institutions.

00:18:18 And look, I'm sympathetic to the instinct. Depending on an American platform for the search layer of a legislature is a real vulnerability, and changing a default is one of the few levers an institution actually controls directly. But there's something telling about the scale of the gesture.

00:18:34 The continent that can't fund its data centers is asserting its independence by changing a setting in a browser. The sovereignty you can reach by swapping a default isn't the sovereignty that matters when the models, the compute, and the capital all sit somewhere else.

00:18:49 I don't say that to mock it. Defaults shape behavior — that's the whole reason the search business is worth what it's worth, and an EU institution moving its traffic to a European engine is a modest redistribution of attention and data. But put it beside the gigafactory story and you see the bind.

00:19:06 Real AI sovereignty requires compute, energy, and capital at a scale Europe is visibly struggling to assemble. Symbolic sovereignty is available off the shelf. The risk is that the second becomes a substitute for the first — that you change the default, declare independence, and keep renting the actual frontier from the same handful of American and Asian firms everyone else rents it from.

00:19:28 The question I'd put to Brussels is the one Rogoff is really asking the whole world: are you building a place in the supply chain, or are you decorating the spot where one was supposed to go?

00:19:39 Let me tie the day together, because the threads are closer than they look. The through-line is early access — who gets to see and use the most powerful models first, and who decides. The executive order spends its energy on that exact question and then hands the answer back to the companies, because the access is voluntary and the licensing is forbidden.

00:19:58 Anthropic answers the question for itself, drawing a security perimeter around NATO and ENISA and a list of allied firms while the government is still writing the benchmark. Microsoft and Mayo answer it for medicine, bundling the most capable clinical model with the most trusted hospital brand.

00:20:14 And Rogoff and the European stories answer it for everyone outside the frame — the countries, the workers, and the institutions reaching for a search engine because the data centers won't come. In every case the capability already exists, and the access is narrowing toward the same small set of hands.

00:20:30 Here's what I'll be tracking. The sixty-day mark on the executive order, when the classified benchmark and the voluntary framework are due — that's when we learn whether this was the first real move toward something binding or a comment period with good intentions.

00:20:44 Whether any major lab actually declines to hand its next model over, now that the order makes participation a choice with no reward attached. Anthropic's prospectus, when the IPO surfaces, and how much of its valuation rests on being the security vendor to the Western alliance.

00:20:59 And whether DeepMind's Co-Scientist leads on liver fibrosis and ALS survive contact with an actual laboratory, or join the growing pile of plausible AI output that nobody verified. One line to leave you with. The government asked, politely, for a look at the frontier.

00:21:14 The labs had already given that look to the militaries. The order didn't create the access — it just discovered, thirty days late, that the access already had an owner. That's the thing I'm watching: not whether the state regulates the frontier, but whether it ever catches up to the fact that the frontier is already governing parts of the world on its own terms.

00:21:33 I'm Jonas.