Model Access Became A Nationality Test

00:00:04 Anthropic senior staff were scheduled to meet Trump administration officials in Washington on Monday, June 15, after the government ordered the company to suspend access to Fable 5 and Mythos 5 on Friday, June 12. That’s the new fact today. On Sunday, the open item was whether the White House or Anthropic would publish a technical basis for the shutdown.

00:00:25 As of the sources I could get today, the answer is still no. We have more process detail, more reported anger, more names in the room, and a sharper labor question. We don’t have a public technical finding that lets outsiders evaluate the risk claim. CNBC says Anthropic received an export-control directive that cited national security authorities and ordered the company to suspend access to the two models for foreign nationals, inside or outside the United States.

00:00:54 Anthropic then disabled the models for all customers to comply. CNBC also reports that the company says it worked with government agencies before release and received approval to deploy the models. According to the same report, the government called Anthropic at one o’clock Eastern on Friday and told the company to disable Fable 5 and Mythos 5 because of an unspecified national security threat.

00:01:18 A formal letter arrived around five-thirty Eastern. That timing changes the institutional picture. A normal product recall gives customers a defect, a mitigation, and a planned return path. This order seems to have given Anthropic a deadline and a category of excluded person.

00:01:35 The Financial Times, as summarized by Techmeme, adds that Anthropic was given ninety minutes to comply and wasn’t given detailed concerns before the export-control order. If that is accurate, the procedure itself becomes part of the risk. You can believe the government had a serious concern and still ask how a customer, a partner, or a foreign-born employee is supposed to reason about a rule that arrives faster than a large company can run an internal incident review.

00:02:03 The reported politics are blunt. Axios quotes an administration official saying, three words, "They screwed us." Axios also reports that Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent on Thursday to raise concerns about security risks in Anthropic’s new models.

00:02:20 Amazon matters here twice. It is a large investor in Anthropic, and it is a cloud and chip partner. So when Amazon raises a concern with senior officials, a major distribution and financing partner is talking to the state from inside the relationship. Anthropic says the concern appears to involve a narrow jailbreak where Fable 5 could be pushed into reading a codebase and fixing software flaws.

00:02:44 The company argues that this standard, if applied across the industry, would halt new deployments for frontier providers. I think that argument has force, but only up to a point. Cyber capability isn’t a normal consumer feature. A model that can materially improve vulnerability discovery and exploit development will draw a different government response than a writing assistant.

00:03:07 Outsiders still can’t see the test, the threshold, or the exception process. The staffing issue may be the bigger change. An X post from prinz quotes the Financial Times saying the AI industry had been working with the U.S. government on keeping foreign national researchers involved in advanced model development, and that the Anthropic directive banned that practice.

00:03:30 I would treat the post as a pointer to FT reporting, not as a final legal interpretation. But the concern is obvious. If frontier model access restrictions reach employees, then AI governance starts allocating immigration status, labor access, export licenses, and company secrecy at the same time.

00:03:48 There is an older template for this in export-controlled defense and semiconductor work. Foreign nationals can sometimes work on restricted technology under licenses, deemed-export reviews, and exception processes. The difference is speed. AI labs rely on global research teams, fast model iteration, cloud access, and remote collaboration.

00:04:09 A rule that appears on Friday afternoon can scramble access, review, patching, evaluation, and the people allowed to explain the model to regulators. So the follow-up from Sunday remains open. Anthropic is in the room with Commerce, the CIA, and White House science officials, according to Axios and CNBC.

00:04:27 The public evidence is still thinner than the public consequence. A written process for technical review, employee access, and restoration would matter more than another anonymous quote from either side.

00:04:40 Nvidia filed on Monday to raise capital in the bond market, and CNBC reports the company is aiming for at least twenty billion dollars in debt. The number may end up closer to twenty-five billion dollars, according to CNBC’s sources. That would be Nvidia’s first bond sale since 2021, before the current AI boom turned the company into the supplier everyone else has to explain in their capital plans.

00:05:04 In 2021, Nvidia raised five billion dollars. In fiscal 2022, it generated about twenty-seven billion dollars in revenue. CNBC says fiscal 2026 sales were two hundred sixteen billion dollars. The filing doesn’t read like a weak company hunting for cash. Nvidia has about seven and a half billion dollars in long-term debt, about one billion dollars in short-term debt, and forty-nine billion dollars in free cash flow in the latest quarter.

00:05:30 The company also raised its dividend in May from a penny per share to twenty-five cents and announced an eighty billion dollar share-repurchase plan. Management has said it plans to return roughly half of free cash flow to shareholders this year. This is a capital-structure story.

00:05:47 Nvidia is borrowing while it is profitable, while investors still want AI exposure, and while the rest of the ecosystem is also tapping markets to finance the buildout. CNBC places Nvidia next to Alphabet, Amazon, and Super Micro in that pattern. Alphabet announced equity-related offerings after more than fifty-five billion dollars of fresh debt since November.

00:06:09 Amazon raised roughly fifty-four billion dollars earlier this year in U.S. and European bond sales and then announced plans for another Canadian debt sale. Super Micro announced seven billion dollars in equity-related financing to cover hardware component purchases.

00:06:25 You can read that as confidence. The AI buildout is large enough that capital markets are funding labs and suppliers. They are also funding cloud providers, server vendors, energy contracts, and the buyback politics around them. Nvidia’s stated use of proceeds is general corporate purposes, including repayment and refinancing of existing debt.

00:06:46 That is deliberately plain language. Companies don’t have to say, in the filing headline, that they are raising the price of admission to the AI economy. The practical effect is that AI compute is turning into a balance-sheet contest. A model company that wants frontier capacity has to find chips, data-center space, power, debt financing, partner guarantees, and a political story that says the whole thing should keep expanding.

00:07:12 Last week we talked about compute as a dependency. Today, Nvidia shows the other side: the dependency has a treasury desk. There is a small dry comedy in the fact that Nvidia is borrowing while also returning cash to shareholders. That is normal corporate finance, and it may be rational.

00:07:29 But it also says something about the distribution of gains. The companies closest to the chip constraint can finance at enormous scale, reward shareholders, and still remain the route through which everyone else buys capacity. The AI economy is rewarding model skill, and it is also rewarding the firms that can borrow, allocate, and price scarce physical capability before the next generation arrives.

00:07:53 I don’t think this bond sale is a turning point by itself. It is a measurement. The market is treating AI buildout as durable enough to underwrite twenty billion dollars or more of new paper. If demand weakens, that debt sits on the company. If demand holds, the companies without access to similar financing will feel the gap first.

00:08:13 The Verge reports that Big Tech lobbyists and the White House are trying to pair federal AI preemption with child online safety legislation. Large technology companies want preemption. A federal AI law with preemption would override a messy state-by-state regulatory system and create one national rulebook.

00:08:31 One federal standard is easier to lobby, easier to staff, and easier to price into product planning than fifty state regimes plus city rules and sector-specific regulators. The vehicle is unusual. According to The Verge, reports leaked that the White House told child safety groups and Big Tech companies it would endorse legislation backed by Senator Marsha Blackburn, the coauthor of the Kids Online Safety Act, as part of an overall AI preemption package.

00:08:58 The politics are messy. House Republicans had just passed their own version of KOSA. Democrats who worked with Blackburn on the Senate version reportedly were not aware their bill might be tied to AI preemption. A separate bipartisan AI preemption bill was already moving around the House.

00:09:15 A Republican lobbyist told The Verge, "No one knows really who’s driving this thing." That sentence is almost too neat, but it captures the policy problem. The same package is trying to satisfy online child safety advocates, Republican populists, Big Tech lobbyists, House leadership, Senate Democrats, the White House, and state officials who don’t want their AI laws erased.

00:09:37 Mike Davis, the Trump-allied lawyer who helped kill a previous AI moratorium in the Senate, told The Verge that any preemption law has to address four constituencies: children, conservatives, creators, and communities. The KOSA attachment seems designed to satisfy the children part.

00:09:53 It doesn’t settle the rest. The Senate version of KOSA includes a duty of care and extends that responsibility to AI companies. The House version weakened that provision, which angered child safety advocates. If Big Tech wants preemption, it may have to accept a stronger child-safety obligation than it would choose on its own.

00:10:12 That trade carries the policy risk. The companies want protection from the states. To get it, they may have to accept federal obligations that touch product design, recommender systems, chatbots, and youth risk. The states, meanwhile, are trying to preserve their ability to regulate discrimination, safety, environmental effects, and consumer protection.

00:10:33 A state law may be clumsy. It may also be the only rule that exists while Congress negotiates. The timing looks bad for the preemption push. The Verge notes that Congress has about a month and a half before a five-week recess and then the general election season.

00:10:48 The calendar also has FISA renewal, immigration, defense spending tied to the war with Iran, a crypto market structure bill, affordability measures, election legislation, and regular budget fights. Even if the White House can force Republican alignment, a Senate path still needs Democrats.

00:11:05 I read the procedural mess as evidence of AI power in Washington. The industry knows state rules are coming. It wants a single national settlement before those rules harden. Child safety may become the price of that settlement, but the package is carrying too many unresolved interests at once.

00:11:22 If it fails this summer, the state-by-state system keeps growing, and every lab, app company, school vendor, and model provider has to plan for a regulatory map that changes by jurisdiction.

00:11:34 Challenger, Gray and Christmas reported that U.S. employers announced ninety-seven thousand six job cuts in May, the highest May total since 2020. The report gives the numbers underneath the AI-layoff argument. The May total was up sixteen percent from April and up three percent from May of last year.

00:11:52 For the first five months of 2026, employers announced three hundred ninety-seven thousand seven hundred fifty-five cuts. That is down forty-three percent from the same period in 2025, when federal workforce reductions distorted the comparison. Against 2024, the year is running roughly even.

00:12:09 The AI-specific numbers are sharper. Challenger says artificial intelligence led cited reasons for cuts for the third month in a row, with thirty-eight thousand five hundred seventy-nine cuts in May. The report says AI "accounted for 40% of all cuts announced in May," up from seven percent in January, twenty-five percent in March, and twenty-six percent in April.

00:12:31 For the year, AI has been cited in eighty-seven thousand seven hundred fourteen cuts, or twenty-two percent of all 2026 layoffs. That already exceeds the fifty-four thousand eight hundred thirty-six cuts attributed to AI in all of 2025. A caveat needs to stay attached to every labor statistic like this.

00:12:50 These are employer-cited reasons in announcements. They aren’t an audit of whether a model replaced a person, whether revenue fell, whether a merger removed a team, or whether management used AI as a tidier explanation for a cost-cutting plan it wanted anyway. Bruce Lambert had the natural reply to Robin Hanson’s post about the figure: what does “accounted for” mean here?

00:13:12 Any serious labor reading has to keep that uncertainty attached. Still, cited reasons have consequences. If executives, investors, and employees keep hearing AI named as the reason for reductions, the labor market will behave as if AI is part of the restructuring force even before economists can isolate causality.

00:13:31 Workers change what they study, managers change what they approve, and recruiters change what they ask for. Boards ask why headcount isn’t falling if other firms say AI is helping them cut. The sector detail matters too. Transportation announced six thousand nine hundred nine cuts in May and more than forty thousand so far this year, up four hundred forty-nine percent from the same period in 2025.

00:13:56 Health care and products manufacturers, including hospitals, have announced more than thirty thousand cuts this year. Pharmaceutical companies announced just over five thousand cuts in May and more than twelve thousand through the first five months, up seven hundred fifty-three percent from the same period last year.

00:14:15 Fintech companies announced five thousand seven hundred thirty-one cuts in May, and the report says most of those cited AI. Hiring doesn’t offset the anxiety. Employers announced eighty thousand four hundred seventy-two planned hires through May, barely above the same point last year and low by pre-pandemic standards.

00:14:34 Technology led May hiring with eleven thousand two hundred fifty announced positions. The labor market isn’t simply shrinking. It is reallocating. The pressure lands on the middle: functions that can be partially automated, teams caught inside mergers, and workers asked to prove they still belong in a workflow whose budget now assumes model assistance.

00:14:55 I wouldn’t turn this into a claim that AI has already taken forty percent of U.S. layoffs. The report doesn’t prove that. It proves that AI is now a common managerial explanation for job cuts, and that explanation is arriving with enough frequency to affect bargaining power.

00:15:12 That is still a serious labor story.

00:15:14 The Office of Management and Budget disclosed three thousand six hundred eleven active or planned federal AI use cases in April, according to Nathan Sanders and Bruce Schneier writing in The Guardian. Their piece is an opinion essay, but the underlying inventory number is the reason it belongs here.

00:15:31 They say the list is up seventy percent from the one published in the final year of the Biden administration. Not every listed use is automatically bad. But the public gets a thin description of systems that may affect grants, prison classification, crisis calls, nuclear reactor response, and government translation.

00:15:50 They walk through several examples. Health and Human Services hired Palantir to scan grant applications for ideological alignment with administration policy. The Federal Bureau of Prisons is developing an AI system to assess the potential for misconduct among newly admitted inmates.

00:16:06 The Department of Veterans Affairs is developing an AI system that listens to calls on the veterans crisis line and gathers outside database information to assess mental state and suicide risk. The Department of Energy is testing AI for nuclear-reactor response.

00:16:21 The State Department ended a program that had used AI to forecast mass civilian killings. You can hear those examples and feel the reflexive alarm. Some deserve it. But Sanders and Schneier draw a distinction that helps. A scary-sounding use case can be a bad policy, a bad model, a bad disclosure, or a plausible system that hasn’t been explained well enough.

00:16:42 Machine translation at Customs and Border Protection is a good example. In some cases, an immediate AI translation tool may be better than no communication at all. In other cases, losing human interpretation can remove context that matters for rights, fear, and coercion.

00:16:58 That is why the disclosure process matters. The authors say the inventory descriptions are usually just a sentence, rarely more than a paragraph. Public consultation is theoretically part of the process, but it rarely happens in a way ordinary people would encounter.

00:17:13 They also say only one of their cited examples proposed public involvement, because the rest were not classified as high-impact use cases. That label appears inconsistently across agencies. The comparison to France and Canada gives us a policy baseline. France’s 2016 Digital Republic Act requires algorithms used to automate administrative decisions to be subject to public records requests, appealable to a human reviewer, and disclosed to the person affected.

00:17:40 Canada has an AI use-case registry and a federal directive requiring risk scoring and impact assessment for automated systems that make administrative decisions about citizens. Sanders and Schneier argue Canada could still improve by requiring public comment and substantive agency responses before sensitive uses go live.

00:17:58 The United States has the inventory. It doesn’t yet have a public process that matches the stakes of the inventory. If an AI system helps decide whether a grant aligns with policy, whether an inmate is routed toward higher security, or whether a veteran in crisis gets escalated, the affected person shouldn’t have to find a GitHub account to know automation is in the room.

00:18:20 This connects back to the private-sector stories, but only lightly. The same institutional problem keeps appearing: AI deployment moves faster than the explanations around it. In companies, that produces pricing confusion, access shocks, and layoffs explained by a phrase in a report.

00:18:36 In government, it can produce automated decisions with public authority behind them. The American state isn’t waiting for a perfect national AI law before using these systems. It is already using them, and the paper trail is too thin for the decisions being contemplated.

00:18:52 A new arXiv paper from Nvidia researchers evaluates a driving vision-language-action model on fifteen thousand nine hundred sixty-eight clip-and-attack pairs. The paper is called Scenario-Specific Safety Envelopes for Driving VLAs. VLA here means vision-language-action.

00:19:08 The evaluated model is Alpamayo R1, a ten billion parameter open-weight driving model. The authors are trying to answer a practical certification question: when does the planner start to fail, and how severe are the failures after that point? Safety policy often compresses that second dimension too much.

00:19:26 A single aggregate score can tell you the model looks safe up to a certain level of noise. The paper says the aggregate threshold is sigma less than or equal to fifty under a fifteen percent average-displacement-error budget. But when the authors break the results out by scenario, four of six well-sampled scenarios tolerate the top of the tested grid, sigma seventy.

00:19:48 Lane keeping and following a vehicle match the aggregate threshold. A small pilot scenario, nudge right, appears tighter, though the authors are explicit that the sample is small. Then comes the result that changes how I’d read the score. The scenarios that tolerate more noise aren’t necessarily the scenarios with mild failures.

00:20:07 The stop-signal and intersection scenarios tolerate sigma seventy, but they also concentrate the largest high-severity shares among well-sampled scenarios. The stop-signal scenario had eleven point eight percent of changed cases in the high-severity bands, and the intersection scenario had eight point eight percent.

00:20:26 Lane keeping, with the tighter sigma fifty threshold, had only two point nine percent high-severity share. In plainer terms: a scenario can look robust by one measure and still carry worse failures when it does break. That is the kind of detail autonomy regulation needs.

00:20:42 If you are approving a driving system, one number per hazard isn’t enough. You need a boundary for when performance degrades and a separate severity profile for what happens after degradation. The authors call that a two-dimensional safety envelope. I would call it common sense made measurable.

00:20:59 Physical AI is full of these hidden compression problems. Reuters, through a Techmeme summary, reported today that Tesla sent Swedish and Dutch regulators Full Self-Driving safety data that traffic-safety researchers called misleading, while the Dutch regulator approved FSD in April.

00:21:16 I haven’t read the full Reuters story, only the Techmeme summary, so I won’t build much on that item. The summary is enough to note the tension: regulators are being asked to judge autonomy systems from company-provided evidence, and outside researchers are questioning how that evidence is presented.

00:21:34 The FactoryLLM paper gives the factory version of the same problem. Its authors built an open-source playground for retrieval-augmented generation in smart factories. They tested three models on thirty maintenance questions drawn from about six hundred pages of cross-machine documentation for an autonomous intelligent vehicle and its mobile planner software.

00:21:55 The system achieved groundedness above zero point eighty-eight, but retrieval precision averaged about zero point forty-eight. The generated answers often stayed grounded in the retrieved text, but the retrieval step pulled too much adjacent noise. That is a useful industrial result because it names the weak link.

00:22:14 In a factory, the model doesn’t only need to sound plausible. It needs to retrieve the right manual page, from the right machine, in the right operational context. A wrong answer can stall a production line or push a technician toward an unsafe repair. FactoryLLM also supports local and open-source models, which matters because factory documentation is sensitive.

00:22:35 The answer can’t depend on uploading proprietary maintenance records to whatever service has the best demo this week. Today’s physical-world AI news is measurement catching up with deployment. Driving systems need scenario-specific safety envelopes. Factory assistants need retrieval metrics that distinguish grounded generation from poor document selection.

00:22:56 Those measurements are the difference between an impressive demo and a system an institution can defend after an incident.

00:23:04 A new medical AI benchmark called ClinHallu contains seven thousand thirty-one validated medical visual-question-answering instances. The paper makes a simple claim: judging only the final answer hides where a medical model failed. A multimodal medical model can misread the image, recall the wrong medical knowledge, or combine correct pieces of evidence into the wrong conclusion.

00:23:25 Those failures can produce the same bad answer, but they demand different fixes and different human review. ClinHallu decomposes each instance into three stages: visual recognition, knowledge recall, and reasoning integration. The authors build the benchmark from four medical visual-question-answering datasets and then use stage-replacement interventions to test what happens when a specific stage is corrected.

00:23:48 They evaluate eleven closed and open models. They report that visual hallucination is generally severe, with average rates above forty percent across the evaluated subsets. VQA-RAD is more visually constrained, while MedXpertQA is more knowledge constrained. PathVQA and MedFrameQA are more balanced.

00:24:05 That level of diagnosis is what medical deployment needs. A hospital doesn’t only need to know that a model gave a wrong answer. It needs to know whether the model failed because it misrecognized a lesion, recalled the wrong association, or made a bad inference after seeing and recalling the correct facts.

00:24:22 Each failure sends the safety work to a different place. Image recognition means data, sensors, and modality coverage. Knowledge recall means training, retrieval, and updating. Reasoning integration means the model may need structured supervision, better checks, or a narrower clinical role.

00:24:38 The paper also reports that trace-supervised fine-tuning helped. Answer-only fine-tuning improved accuracy but gave limited stage-wise gains. Full trace supervision produced the highest accuracy and lowest hallucination rates. That isn’t a deployment guarantee.

00:24:53 The authors say the benchmark focuses on visual-question-answering tasks and doesn’t yet cover long-form report generation or real clinical decision-support workflows. Still, it gives hospitals and regulators a better question than, did the model get the answer right on average?

00:25:09 Ethan Mollick had a related public-policy note today. He argued that AI has reached a level where universal tutors, co-scientist and replication systems, and remote medical help could produce large social benefits, but require public research, consensus, and transparency.

00:25:24 He added that current open models are good enough for some of these projects when scaffolded properly, while co-scientist systems still benefit from frontier models. I like that formulation because it doesn’t pretend private capability automatically becomes public value.

00:25:39 Remote medical help is a perfect example. If a model can answer triage questions in places without enough clinicians, that could matter enormously. If it hallucinates visually, recalls the wrong fact, or gives a confident answer where referral is needed, the same system can move risk onto the person with the least institutional protection.

00:25:58 Public research matters because the necessary evidence isn’t only model accuracy. It is who gets helped, who gets misled, who can appeal, and who pays when the system is wrong. Today’s episode started with the state deciding who can touch a frontier model. It ends with the institutions that would like to use models in factories, cars, agencies, and clinics trying to explain what happens when the system makes contact with the world.

00:26:23 The next useful evidence is a written Commerce process for foreign researchers and model access. Jonas.