Power Started Asking For Receipts

00:00:04 Anthropic is now trying to turn the Mythos and Fable dispute into a repair process. The New York Post reported Thursday that Anthropic executives have floated a proposal to Commerce Secretary Howard Lutnick. The proposal would put the company into closer working contact with the White House and commit it to faster remediation when officials raise security concerns about its models.

00:00:28 The reported proposal doesn't restore access by itself. The same story says the talks are progressing, and that the exact timetable for a permanent fix remains unclear. That matters because the question we carried out of Wednesday wasn't whether Anthropic could win the public argument.

00:00:47 It was whether a written restoration path would appear after the government told a private AI lab to restrict access to two frontier models, and the lab responded by taking them offline. On Thursday, the answer is still incomplete, but the negotiation has a more definite outline.

00:01:05 The Post says Tom Brown, Anthropic's cofounder and chief compute officer, and Sarah Heck, its head of public policy, are taking leading roles in the talks. It also quotes an Anthropic spokesperson saying, "Both parties are working quickly to get this resolved," and describing the company as working with the administration toward shared goals around critical infrastructure and cyber defense.

00:01:30 WIRED gives the sharper backstory. Its report says the administration's action followed concern about SK Telecom's access to Claude Mythos, plus Amazon researchers flagging alleged vulnerabilities in Fable 5. U.S. officials were worried about what they alleged were SK Telecom's ties to China.

00:01:49 Anthropic had given SK Telecom access through Project Glasswing, a controlled access program for a small group of trusted organizations. WIRED reports that SK Telecom was one of roughly 150 companies in the expanded program, alongside Samsung Electronics and the Korea Internet and Security Agency.

00:02:08 The ugly operational detail is nationality. WIRED reports that the White House ordered Anthropic to revoke access to Mythos and Fable 5 for all foreign nationals, including immigrants inside the United States. Rather than build an access system around nationality while preserving privacy, Anthropic disabled the models entirely.

00:02:30 A model safety dispute became a citizenship-and-access problem inside a lab product surface, and the fallback was to remove the tool from everyone. There is also an investment and alliance layer here. SK Telecom invested 100 million dollars in Anthropic in 2023, tied to a commercial partnership for a telecom-specific AI model.

00:02:51 WIRED says the company generated only about 1.9 million dollars in China revenue in 2024 and employed seven people there. The report also walks through older China Unicom ties and broader SK Group business interests. Export-control stories get messy exactly this way.

00:03:08 The same company can be a South Korean partner, a U.S. investor relation, a telecom operator, a historical China-linked business actor, and a participant in a cyber-capability access program. My read is that Anthropic's offer is less about apologizing and more about making government trust legible.

00:03:27 Faster communication, named remediation procedures, and severity scoring let the administration say it has control without writing a full statute first. That may get Mythos and Fable back online. It also means the lab's most sensitive models now live inside a standing political relationship, not just a product release process.

00:03:48 The missing piece is still the written rule. Who can lose access? Who can appeal? What evidence triggers a shutdown? How does a company tell affected customers without exposing private security findings? Until those answers are public, model access remains negotiable power with a support ticket attached.

00:04:07 Axios published two useful pieces Thursday that make the Anthropic fight look less like an isolated dispute and more like the working method of U.S. AI policy. Ashley Gold's piece describes what she calls the Trump administration's shadow AI policy. The administration came in arguing that AI shouldn't be regulated in the heavy European style, and it rescinded Biden-era requirements.

00:04:32 But according to Axios, the practical result hasn't been a blank space. It has been company-specific intervention. It has also included voluntary frameworks, executive orders, procurement rules, export controls, and case-by-case pressure. The article puts it plainly: "Without clear rules, companies can find themselves navigating personalities and broader politics as much as policy."

00:04:59 A lab doesn't just ask whether a model meets its safety case. It asks which official is angry, which cabinet secretary signed the letter, which competitor raised the concern, and which agency claims jurisdiction. It also asks which summit is happening while the dispute is open.

00:05:17 That isn't a stable operating environment, even if you like the policy outcome in a particular case. Maria Curi's companion piece names the people. Commerce Secretary Howard Lutnick is central because his signature was on the letter that imposed export controls on Anthropic and created, in Axios' wording, a licensing regime that could eventually affect other AI labs.

00:05:41 Chris Fall at Commerce's Center for AI Standards and Innovation has been holding technical meetings in Washington. Treasury Secretary Scott Bessent enters through the financial-sector concern about Anthropic's Mythos, even though Treasury doesn't directly oversee AI testing.

00:05:58 White House chief of staff Susie Wiles is described as receptive to Bessent's concerns and as helping reopen communication with Anthropic. Ryan Baasch at the National Economic Council is carrying policy work associated with David Sacks and Sriram Krishnan. National Cyber Director Sean Cairncross is also holding technical meetings and reportedly clashing with the Treasury side over urgency.

00:06:24 That is a lot of names, but the names are the point. When Congress hasn't produced a settled AI law, and the administration doesn't want to call its actions regulation, authority travels through people. It travels through letters, procurement language, export controls, voluntary testing frameworks, and private meetings.

00:06:44 That can move quickly. It can also make the industry exquisitely sensitive to personnel changes. Axios quotes a White House spokesperson saying, "The President's team, including Secretary Bessent and Sean Cairncross, is working closely together to strengthen America's cyber and national security, protect critical infrastructure, and ensure the United States remains the global leader in AI innovation." Fine.

00:07:11 That is the public line. The reported internal friction still matters because the United States is home to the most important frontier labs. The people now improvising this regime are also sending signals to Europe, South Korea, the Gulf states, Japan, and every buyer who needs to decide whether a U.S.

00:07:30 model can be part of critical infrastructure. Dean Ball also announced Thursday that he is joining OpenAI on July 6 to lead a new Strategic Futures team with a mandate to help company leadership shape frontier AI policy. I won't make that more dramatic than it is.

00:07:47 One person moving jobs is one person moving jobs. But the direction of travel is visible: policy people are becoming core lab operators, and lab operators are becoming policy actors. AI governance is being made before it is being named. A hospital, a bank, a defense contractor, a school system, or a foreign telecom firm may never see a bill in Congress that tells it the rule.

00:08:12 The first rule may arrive as an access decision, a contract clause, a model-review framework, or an export-control letter.

00:08:20 The Federal Energy Regulatory Commission moved Thursday on large-load interconnection, and NVIDIA immediately presented it as an AI infrastructure decision. The NVIDIA blog post is corporate advocacy, so read it with that in mind. It is still useful because the company says the operational part plainly.

00:08:39 Large AI loads, semiconductor facilities, and advanced manufacturing sites need to connect to the electric grid. Those connections can take years. The new framework, in NVIDIA's telling, pushes large customers to become active participants rather than passive entrants in an overloaded queue.

00:08:57 That means paying for network upgrades, bringing new generation online, and offering flexible load that grid operators can curtail or shift when the system is stressed. One detail is concrete enough to matter. NVIDIA says customers that can demonstrate flexibility may move through the process on accelerated timelines, with study periods potentially as short as 60 days under Energy Secretary Chris Wright's directive.

00:09:24 Sixty days isn't a data-center buildout. It is a study window. But compared with multi-year interconnection delays, the policy signal is easy to read: if a large load can behave like a grid asset instead of a pure demand spike, regulators may let it move faster.

00:09:41 NVIDIA also makes the ratepayer argument. Electric grids have high fixed costs. Add demand efficiently, and those costs can be spread over a broader base. The post cites Lawrence Berkeley National Laboratory for the claim that every 10 percent increase in state electricity consumption correlates with about a six-cent-per-kilowatt-hour reduction in retail prices.

00:10:04 It also cites North Dakota, Mississippi, Louisiana, Virginia, and a PG&E forecast that each new gigawatt of data-center load could reduce electric rates by one to two percent under the right conditions. That is the pro-growth case. The consumer-risk case came from Axios' Arizona reporting, and it is much harder to hand-wave away.

00:10:25 Kevin Thompson, a member of the Arizona Corporation Commission, said at an Axios event in Phoenix that what took utilities more than 100 years to build may need to be doubled within four to five years to keep up with demand. Then he asked the obvious question: how do you do that without putting the cost onto existing customers?

00:10:46 Arizona is useful because it makes the trade-off physical. The Phoenix area is hotter, water is scarce, and the data-center campus isn't a slide in an investor deck. Axios reports that Arizona paused certain data-center tax incentives for three years. Texas Governor Greg Abbott has called for new rules that would remove some data-center tax incentives.

00:11:09 Ohio Governor Mike DeWine paused state sales-tax breaks for data centers. Republican-led and Democratic-led jurisdictions are looking at the same pressure: if AI companies want power and water, somebody has to decide which public benefits justify the concessions.

00:11:26 The cooling detail is almost too perfect. Google's first Arizona data center uses air-cooled technologies instead of more water-intensive evaporative cooling. The site is about 30 miles east of downtown Phoenix. That saves water but generally uses more electricity.

00:11:42 Google says its Arizona site recirculates water in a closed-loop system for chips while relying on air-based chillers for the facility. Thompson says the company originally proposed evaporative cooling before changing course. A data center in Arizona is a water decision, a power-price decision, a worker heat-exposure decision, and a tax-incentive decision.

00:12:05 Thompson's closing line to Axios should discipline this whole debate: residents want their air conditioning running when it is 120 degrees outside, and they want it to be affordable. So when NVIDIA welcomes FERC action, I understand why. The AI industry can't scale on permission queues that were built for a different demand curve.

00:12:26 But the public bargain has to be written in dollars, megawatts, gallons, outage risk, and local tax policy. If the AI factory is flexible, prove it in the tariff. If it reduces household bills, show the bill. If it needs a faster path to the grid, say who pays for the road.

00:12:44 OpenAI and Boston Children's Hospital reported Thursday that an OpenAI reasoning model helped clinicians revisit 376 previously unsolved rare-disease cases and clarify 18 diagnoses. That is a strong claim, and it is narrow enough to talk about without turning it into a miracle story.

00:13:03 NBC News reported that the work was announced in NEJM AI and came from Boston Children's Hospital's Manton Center for Orphan Disease Research in collaboration with OpenAI. The patients had already gone through genetic testing and expert review. The research team fed the o3 system clinical notes, symptom descriptions, and filtered lists of genes that might explain the patients' conditions.

00:13:28 Human researchers reviewed the system's outputs before any final diagnosis. Catherine Brownstein, scientific director of the genetic investigations arm of the Manton Center, gave NBC the line that sets the altitude correctly: "It got almost 5% new diagnoses, which doesn't sound like a lot, but considering how many times these had already been analyzed, that's a huge number, and each one means an answer for a family." That is the right way to hear the number.

00:13:59 Five percent would be weak if this were first-pass triage across easy cases. In a backlog of already-analyzed rare cases, five percent is meaningful. NBC walks through one human case: Kyra Benton. She began moving differently from her peers at nine, walked on tiptoes, struggled to run normally, saw specialists, had worsening health, developed severe heart problems, underwent a tracheotomy at 13, and came to terms with never knowing the cause.

00:14:28 Last year, researchers called before her 20th birthday with an answer: myofibrillar myopathy, a progressive genetic neuromuscular disorder. The science problem here isn't that no one cares. A human genome has around 20,000 protein-coding genes. Medical knowledge changes constantly.

00:14:46 A case that was unsolved when first reviewed may become solvable after a later paper links a gene to a disease. Suyash Shringarpure at OpenAI put it in operational terms to NBC: "Maybe a case remained unsolved when it came to them first, but a year later a paper was published that clarifies the link between the gene and the disease." A large language model isn't being asked to be a doctor in a room with a child.

00:15:14 It is being asked to connect clinical features, inheritance patterns, variant data, and scattered literature in a workflow where specialists are time-limited. There are limits. NBC notes that seven of the identified diagnoses were rediscoveries, meaning a treatment team elsewhere had already identified a diagnosis but it had not been shared with researchers around the world.

00:15:39 That sounds less impressive until you think about what a diagnosis is for. Brownstein said rediscoveries still matter because, when new treatments come online, researchers can find patients quickly and put them in line for new therapies. The OpenAI health video adds the labor model around the product.

00:15:58 A pediatrician says she has practiced for 16 years and leads a team of hundreds of physicians evaluating OpenAI's health models. The physicians look at accuracy and at the impact a response may have on a person. One line from the video is worth keeping: "How you say something often matters as much as what you say." That isn't a benchmark claim.

00:16:21 It is a medical deployment claim. Tone can change whether a person panics, delays care, misunderstands risk, or asks a better question at the clinic. OpenAI says every week more than 230 million people turn to ChatGPT for health and wellness questions. If that number is even directionally right, the health interface is already a public institution of a sort, whether regulators have caught up or not.

00:16:48 A practicing physician in the video says she can see about 15 patients a day in clinic, while this work may affect millions of lives. That is both the promise and the anxiety in one sentence. My caution is simple. The rare-disease study is a clinician-mediated workflow with expert review.

00:17:07 The consumer health product is a mass conversation surface. They overlap, but they aren't the same. The Boston Children's work shows a reasoning model helping specialists reopen hard cases. The health video shows OpenAI building a physician review apparatus around everyday answers.

00:17:25 The first deserves attention because it is specific. The second deserves scrutiny because it is huge.

00:17:32 Bloomberg, via Techmeme, reported Thursday that the European Union may say as early as next week that Amazon Web Services and Microsoft Azure appear to meet the criteria for regulation under the Digital Markets Act. There isn't enough public detail yet to do more than treat this as a pending regulatory signal.

00:17:52 But even the preliminary finding would matter because it pulls cloud infrastructure toward the same competition-rule machinery Europe has used for dominant digital platforms. That matters for AI because frontier capability isn't only model weights and research talent.

00:18:10 It is also cloud contracts, graphics processor access, data gravity, egress fees, procurement channels, and the ability to move workloads without being punished for leaving. Europe's technology sovereignty argument often sounds abstract until you put a cloud bill next to it.

00:18:29 A government may want local models, local data rules, and local procurement conditions. But if the practical compute path still runs through AWS and Azure, the cloud providers have leverage over the timeline, the cost, and the design of any local AI stack. Digital Markets Act scrutiny wouldn't magically create European compute capacity.

00:18:51 It could, however, change the bargaining terms around interoperability, bundling, self-preference, data portability, and contractual lock-in. This also sits next to the U.S. policy story from earlier in the episode. The United States is improvising model access rules through export controls, voluntary frameworks, and procurement decisions.

00:19:14 Europe is more likely to formalize market-power rules and then apply them to infrastructure layers. Neither path is neutral. One path gives executive officials and agencies more room to act case by case. The other path gives regulators a durable legal instrument, but it can move slowly and collide with commercial cloud architecture that wasn't designed around public portability.

00:19:39 For AI labs, cloud dependence is governance dependence. OpenAI's Microsoft relationship, Anthropic's Amazon relationship, and every enterprise's AWS or Azure deal are more than vendor arrangements. They affect who can scale, who can switch providers, who gets favorable pricing, who sees sensitive workloads, and who becomes hard to dislodge once the operational habit forms.

00:20:04 If the EU does unveil preliminary findings next week, the first practical detail will be the scope. Does the case treat cloud as a general platform market, or does it name AI compute and enterprise AI workloads directly? The second detail will be remedies. A designation without meaningful portability or contracting changes would be mostly symbolic.

00:20:27 A designation with teeth could make cloud infrastructure part of the open-versus-closed AI power fight in a way that model announcements alone don't.

00:20:37 Waymo recalled nearly four thousand robotaxis to restrict them from highways while it works on behavior around construction zones, according to TechCrunch reporting summarized by Techmeme. This is a smaller story than the policy and medicine items, but it belongs in the episode because it is physical AI meeting ordinary disorder.

00:20:58 The filing reportedly cited more than 13 instances of cars driving into highway sections under construction. That isn't a philosophical debate about autonomy. It is a road crew, temporary barriers, changed lane markings, unfamiliar signs, and a vehicle making the wrong call around people who didn't consent to be part of a model evaluation.

00:21:20 The number is also interesting in proportion. Four thousand vehicles is a fleet. More than 13 instances isn't a mass casualty pattern, at least from the information available here, but it is enough to trigger a recall and a highway restriction. That is the regulatory bargain for physical deployment: when the system moves through public space, edge cases become public events.

00:21:44 You can't keep them inside a dashboard. Construction zones are hard for human drivers too. That isn't an excuse; it is the point. Human drivers use social cues, worker gestures, local improvisation, and a lot of fragile judgment when the road stops looking like the map.

00:22:00 Autonomous systems need to handle that reality before highway expansion becomes a normal service promise. If they can't, the company can still operate in narrower domains, but the public should hear the boundary clearly. There is a broader lesson here for AI deployment outside screens.

00:22:18 Software companies often prefer gradual rollout language: limited availability, staged expansion, internal testing, and measured release. Roads don't care about release language. A highway lane either has workers in it or it doesn't. A barrier is either interpreted correctly or it isn't.

00:22:37 A construction zone in the rain at night near a confused human driver isn't a benchmark; it is the product environment. Waymo deserves credit for restriction if the recall removes a known unsafe behavior from public roads while the company fixes it. The next evidence should be literal and dull: filing language, incident counts, the software remedy, the affected operating domains, and whether regulators accept the fix.

00:23:03 If the autonomy industry wants trust, it has to keep making its boundaries visible when the system is embarrassed.

00:23:10 Put the day together and the shared word is permission. Anthropic is negotiating permission to restore two powerful models after a government access fight. The White House is building AI policy through officials, letters, frameworks, and procurement decisions before Congress has written a stable national rule.

00:23:29 FERC is changing the permission path for large electrical loads, while Arizona is asking whether data centers deserve tax incentives, water trade-offs, and faster utility expansion. OpenAI and Boston Children's are testing a medical version of the same question: can a reasoning model reopen cases that specialists already reviewed, and under what human supervision?

00:23:52 Europe may pull AWS and Azure toward Digital Markets Act regulation, which is another way of asking whether cloud giants have too much power to set the terms of AI infrastructure. Waymo is being reminded that public roads grant permission one domain at a time. I don't think that makes the day a grand pivot.

00:24:10 Most of these are partial moves. Some are reported negotiations, preliminary findings, company posts, and recalls. But the direction is consistent enough to matter: AI power is being translated into institutional permission systems. A model can be good and still need a regulator.

00:24:27 A data center can be profitable and still need a grid study. A diagnosis tool can be useful and still need a clinician. A robotaxi can drive well most of the time and still lose highway access because construction zones are part of the world. The open question from earlier this week remains open in a narrower, more useful way.

00:24:47 Anthropic and the administration may resolve Mythos and Fable soon. If they do, the restoration process will tell us more than the press quotes. A public written process would mean model access is becoming governed infrastructure. A private handshake would mean access is still being handled as relationship management at the edge of national security.

00:25:08 Anthropic's written restoration process, if it appears, will show whether this becomes a rule or stays a relationship. Jonas.