The Access Deal Moved To The Summit Table

00:00:04 On Wednesday at the G7 summit in Evian-les-Bains, France, the CEOs of Anthropic and Google DeepMind asked world leaders to put the United States at the center of global AI rules. CNBC reports that Dario Amodei and Demis Hassabis made the case during a closed-door lunch with G7 leaders, President Trump, Sam Altman, and about a dozen tech executives in the room.

00:00:27 Canadian prime minister Mark Carney, according to CNBC's sources, agreed that the U.S. could lead an AI coalition. That meeting would matter if it only concerned standards. The timing made it sharper. Anthropic's Fable 5 and Mythos 5 models had just been blocked by the U.S.

00:00:44 government on national security grounds, after Amazon reportedly warned the White House that certain safety checks could be bypassed. Anthropic shut the models down altogether. So the summit conversation happened while American model access had stopped being only a product question.

00:01:02 It had become a diplomatic question. CNBC says Amodei told the group that cooperation should include structured access to frontier models and trade in chips and critical components in a way that excludes China. OpenAI circulated language calling for an international forum that could set testing standards, provide impartial analysis of capabilities and risks, and give countries a place to cooperate.

00:01:27 That's the polite version of the argument: if these systems are important enough to restrict, they're important enough to govern through something more formal than a sudden shutdown and a few employee quotes. My read: the lab CEOs were trying to convert a crisis into a regime they can live inside.

00:01:46 That's how powerful firms behave when a state discovers it can interrupt their distribution. They ask for rules, recognized testing, and a club whose members can keep buying frontier access. The same club would also decide who stays outside it. If access to models and chips is structured to exclude China, the U.S.-led coalition becomes more than a safety forum.

00:02:09 It becomes an export-control architecture with a softer tablecloth. This partly pays off one of Tuesday's open questions. We still don't have a public written restoration process for Fable and Mythos. CNBC says Anthropic remains in negotiations with the Trump administration.

00:02:26 The fact that this moved to a G7 lunch means the company has the ear of heads of government. It doesn't yet mean a developer, hospital, bank, or ministry outside the United States knows what happens when its model access disappears. One sentence shouldn't get lost under the summit choreography: access has become a state-mediated privilege.

00:02:48 Maybe that's justified for models that can automate cyber operations or support biological misuse. Maybe the current action was too broad or too improvised. We don't know enough from the public record to settle that. But countries are now being asked to accept American leadership over the very system that just proved America can turn the tap off.

00:03:10 TechCrunch's Rebecca Bellan reports the objection more directly. French president Emmanuel Macron warned the G7 lunch that if the U.S. can, quote, "turn off the switch," it could damage both European customers and the American AI companies selling to them. Indian prime minister Narendra Modi also raised concern about the Anthropic block, according to Financial Times reporting summarized by TechCrunch and Techmeme, saying democratic nations need access to top models to protect critical infrastructure.

00:03:42 This wasn't anti-American posturing. Macron and Modi weren't saying their countries don't want American AI. The article's headline gives the blunt version: they want American AI, but they don't want American interruption. The commercial risk and the sovereignty risk have merged.

00:03:59 A startup in Paris, a ministry in New Delhi, or a utility operator in another allied country can build on a frontier model and still discover that the dependency sits in Washington. TechCrunch also reports that G7 leaders discussed a trusted-partners scheme. The idea would be to grant access to advanced models from firms like Anthropic and OpenAI to countries and companies that use those systems to strengthen defenses against rivals such as China.

00:04:28 In theory, that gives allies an answer: access wouldn't be open to everyone, but trusted governments and trusted firms could understand the path in. In practice, a scheme like that needs answers before it becomes more than a diplomatic placeholder. Who decides which firms count?

00:04:45 Do reviewers look at shareholders and customers, or at the workload itself? Does incorporation matter? Do employees' nationalities matter? And when access is denied, does anyone receive technical evidence, or only a national-security conclusion? Aidan Gomez from Cohere gave TechCrunch the Canadian-enterprise-AI version of the concern.

00:05:06 He said dependence on a small handful of big tech companies is dangerous to resilience. I would separate the sales pitch from the underlying point. Cohere has an obvious interest in arguing for alternatives to U.S. hyperscaler-controlled frontier models. But the last week gave that argument a useful exhibit.

00:05:26 A model can be technically excellent, contractually available, and operationally gone after a government decision. There is also an uncomfortable institutional asymmetry here. American labs want international standards because fragmented rules are expensive and slow.

00:05:42 Allied governments want reliable access because fragmented access is risky and slow. Those interests overlap without becoming identical. The labs want a system that protects their license to operate and keeps their products usable abroad. Governments want assurance that the model layer won't become a diplomatic pressure point every time Washington changes its risk calculation.

00:06:06 The missing document is still a written access process. The strongest possible version would say which capabilities trigger review, what evidence threshold applies, which users are affected, how appeals work, how restoration works, and how emergency restrictions expire or become permanent.

00:06:25 Without that, trusted-partner language can sound reassuring while leaving customers with the same old dependency, just dressed in summit language.

00:06:34 Stuart Russell's Guardian column on Wednesday is opinion, and it should be treated as opinion. It still names the safety argument underneath the politics. Russell connects three events: Anthropic's recent discussion of recursive self-improvement, the Fable and Mythos export-control fight, and the fear that increasingly capable models could automate cyber operations at a level governments can't tolerate.

00:06:58 Russell says Anthropic had described early signs of recursive self-improvement, meaning a system devises ways to improve its own intelligence and then uses the improved system to continue the cycle. He also says the latest Anthropic model showed the ability to conduct end-to-end cyber attacks without human assistance.

00:07:16 That second claim depends on Anthropic's internal evidence, and the government's public explanation remains thin. I would keep the skepticism alive there. We have a policy action, an affected company, outside experts arguing comparable capabilities exist elsewhere, and not enough public technical detail to evaluate the boundary.

00:07:36 Still, Russell's regulatory conclusion isn't exotic. He argues for a licensing regime that requires a minimum safety standard before systems are built and released. He compares that to nuclear power, airplanes, buildings, elevators, and even hairdressers and sandwich makers, which is a dry list but a useful one.

00:07:54 He isn't saying AI is exactly like those industries. He is saying society already licenses dangerous systems before the accident, instead of discovering afterward that voluntary testing wasn't enough. The hard part is that AI risk now travels through ordinary product distribution.

00:08:11 A model is released through an API. A customer upgrades a workflow. A security team gets a more capable assistant. A foreign subsidiary loses access. A government discovers that the same capability can help defenders and attackers. Nobody has to roll a machine into a town square for the state to care.

00:08:28 The risk travels through credentials, cloud regions, employee nationality rules, model-routing contracts, and logs that may or may not be available when something goes wrong. That's why the current U.S. action is so unsatisfying even if you accept the premise that powerful cyber capabilities need review.

00:08:46 A licensing system would have to be published, durable, and technically legible enough for serious users to plan around. An emergency block can be defensible in an actual emergency. Emergency governance becomes corrosive when every affected party has to infer the rule from the interruption.

00:09:03 I don't think Russell's column settles the policy. It raises the bar for the policy. If governments believe frontier models can create cyber risk on the scale he describes, the answer can't be a one-off freeze that catches foreign nationals, allied customers, and the company itself in the same net.

00:09:20 The answer has to be a testable standard, a review body, an incident process, and a restoration path. Otherwise, the public gets the fear without the discipline.

00:09:30 The European Commission published its 2026 State of the Digital Decade package on Wednesday, and the timing was almost too neat. While Macron was warning that the United States could switch off model access, Brussels was putting numbers under the anxiety. The Commission says Europe has made progress on its 2030 digital targets, including connectivity, business digitalization, common infrastructure, and public services.

00:09:56 Then it names the gaps. Europe still trails on foundational technologies and computing capacity. It also names cybersecurity, advanced digital adoption, digital skills, and the ability to scale companies. The public-opinion numbers are sharper. In the Special Eurobarometer survey, 82 percent of European citizens said the EU should reduce dependency on digital technologies from third countries.

00:10:21 Eighty-five percent said the EU should prioritize investment in digital services developed and controlled in Europe. Fifty-eight percent said they would switch to an EU-based digital provider even if it cost slightly more. And 87 percent said online manipulation, including foreign interference, AI-generated content, and deepfakes, threatens democratic processes.

00:10:44 Those figures measure legitimacy rather than model-card performance. Europe can spend years arguing about whether it has the capital, energy, talent, and political will to build frontier systems that compete with the American labs. The public is already telling Brussels that dependency itself has become politically expensive.

00:11:04 The G7 fight over model access gives that sentiment a concrete object. It is easier to sell digital sovereignty when a major U.S. model supplier has just been forced to freeze its most capable systems for foreign users. The digital health report shows why this gets more serious outside consumer software.

00:11:23 The EU-27 composite score reached 87 percent, up four percentage points from 2024. Twenty-three member states report that most of their national population is technically able to access electronic health records through the national service. But the report also says medical images are only at 35 percent maturity, and medical devices and implants are at 59 percent.

00:11:46 Public health providers are better connected than private providers. Mental health facilities, rehabilitation centers, and geriatric nursing homes are less connected than other care settings. So when people talk about AI in health care, the constraint isn't only whether the model can summarize a chart.

00:12:05 It is whether the data is available, interoperable, governed, and connected to the care setting where the decision happens. The European Health Data Space starts applying from March 2027, and the Commission expects it to clarify rights and obligations for health data use.

00:12:22 That is the slow institutional base under the faster model debate. My opinion is that Europe has a harder task than the American labs do this week. The labs need a club that preserves access and manages risk. Europe needs a public case for why autonomy is more than industrial nostalgia.

00:12:40 The Commission's own reports help make that case, but they also expose how much work remains. Computing capacity, cybersecurity, skills, scale-up finance, and health-data interoperability are all gaps at once. Sovereignty can't be bought with one national model contract.

00:12:57 It takes procurement, infrastructure, regulation, and coordination, and every missing piece has a billing address.

00:13:04 The other AI power story on Wednesday was less dramatic than the G7 lunch, but it may be just as revealing. Techmeme summarized Financial Times reporting that Odyssey, a world-model startup, raised 310 million dollars from Amazon and others at a 1.45 billion dollar valuation.

00:13:22 The same item says Odyssey will use AWS as its preferred cloud partner and deploy Amazon's Trainium chips. That is a financing item, and it should be treated at that altitude. It doesn't prove AWS has won the accelerator market, and it doesn't prove world models are about to work at industrial scale.

00:13:42 It does show how cloud providers are using capital, distribution, and custom silicon together. If Amazon invests in the company, becomes the preferred cloud, and gets Trainium into the workload, the startup's model roadmap and Amazon's chip strategy begin to reinforce each other.

00:14:00 The investment is more than money. It is a demand-generation instrument for infrastructure Amazon wants the market to take seriously. Two other items point in the same direction from different angles. Techmeme summarized Nikkei Asia reporting that BYD, Google, AMD, Tesla, and others are increasingly looking to Samsung for advanced contract chipmaking as TSMC capacity remains constrained by AI demand.

00:14:27 The exact customer negotiations are reported by Nikkei, so I wouldn't overstate them from an aggregator blurb. But the basic pressure is familiar: when the premier foundry is full, customers start treating second-source capacity as a board-level supply concern, even when it comes with yield, packaging, and ecosystem questions.

00:14:48 Then there is Kazakhstan. Techmeme summarized Bloomberg reporting that Kazakhstan signed accords with Firebird on Nvidia-related computing projects that could bring up to 10 billion dollars in investment, as the country tries to build what it calls a Data Center Valley.

00:15:06 That phrase has the slightly promotional smell every national industrial cluster gets when it is still more aspiration than reality. But the direction is clear enough. Countries that aren't frontier-model headquarters are trying to become compute geographies. They can offer land, energy, regulation, and political attention.

00:15:27 The prize is more than data-center rent. It is a claim on the next layer of industrial capability. Put those together and you get a less tidy version of AI supply. One company gets cloud financing tied to a specific chip stack. Large customers look for foundry alternatives because the dominant supplier is constrained.

00:15:48 A state tries to turn energy and territory into Nvidia-linked investment. This side of AI competition looks less like software and more like aviation, semiconductors, or liquefied natural gas. It runs through capital commitments, bottleneck suppliers, long lead times, and national development plans.

00:16:08 I think this is where the access-control debate and the compute debate meet. A trusted-partners model for frontier systems can't be only about who gets the API. It will have to care about chips, datacenter leases, export approvals, and cloud credits that make training or inference economically possible.

00:16:28 If the G7 coalition becomes real, the paperwork will probably say safety and standards. The bargaining underneath will be about capacity.

00:16:37 A new complaint in the Northern District of California, Gagleard v. Perplexity AI, is a smaller item than the G7 summit, but it sits on the same map of control. CourtListener's metadata for the filing says Perplexity's search box returns written responses with inline citations and that, to produce those responses, Perplexity leverages leading large language models including OpenAI's ChatGPT, Google's Gemini, Anthropic's Claude, Moonshot's Kimi, and Nvidia's Nemotron.

00:17:07 That is an allegation in a complaint, not a finding by a court. The article extractor hit a 403 on the court page, so I am not going to pretend I have read the full document beyond the metadata the Braid source tool returned. Still, even the excerpt captures an important legal structure.

00:17:26 The user sees one answer engine. Behind it sits a stack of model providers, citation sources, subscriptions, routing decisions, and product claims. When a lawsuit names the answer engine, the liability story has to decide how much of that stack matters. This is going to keep happening.

00:17:45 AI search products made a promise to users: ask naturally, receive a synthesized answer, and follow the citations if you need to. Publishers, researchers, and businesses hear a different promise. Their work can become part of a response layer that may or may not send traffic, payment, or credit back in a way they recognize.

00:18:06 Courts are where that mismatch becomes a list of claims. The interesting part for IMPULSE is not whether Perplexity wins or loses this specific filing. We are too early for that, and a complaint is designed to tell one side's story. The institutional question is how the law assigns responsibility when an AI interface depends on other labs' models.

00:18:28 If Perplexity uses several model providers, and a user receives an answer that allegedly infringes, misattributes, or damages someone, where does the duty sit? With the product that displayed the answer? With the model provider whose system generated some part of it?

00:18:45 With the source website whose material was summarized? With the customer who prompted it? Companies would very much like these boundaries to be unambiguous, because that makes insurance, contracts, and product launches easier. The product experience is going the other way.

00:19:03 Users don't experience five providers. They experience one box. That box increasingly sits between the public and the web, official documents, commercial knowledge, and medical or legal information. This is another reason the G7 conversation matters beyond geopolitics.

00:19:20 If frontier access becomes structured by trusted partners, testing forums, and export permissions, the answer-engine layer may inherit those divisions too. A search product's capabilities could vary by jurisdiction, model provider, content license, and government approval.

00:19:38 That changes who gets a high-quality answer, who gets a degraded answer, and who can audit the chain that produced it.

00:19:46 The most useful counterweight to summit politics today came from the science and physical-deployment items. Latent Space published a long interview with Joseph Krause of Radical AI about self-driving labs for materials. His point was refreshingly physical: AI for materials isn't like representing a small molecule as a string.

00:20:06 For alloys, performance depends on microstructure and synthesis. It also depends on homogeneity, phase distribution, processing, and manufacturing. He said, quote, "there is no one model that can oneshot a new material." Krause says Radical has made about 1,200 alloys in five or six months, including roughly 300 that are new to the literature and about 10 that look especially promising.

00:20:34 That is interesting. But he also talks about 15-to-30-year materials timelines and a qualification process for aerospace or defense alloys that can take around a decade. A model can propose a composition. It can't wish away tensile testing, corrosion behavior, supply-chain constraints, manufacturability, or the FAA's need to know that a turbine material won't betray a passenger aircraft.

00:20:58 The papers in today's feed had the same temperament when they were strongest. A clinical decision-support paper proposes a system using treatment-effect estimation, patient digital twins, reinforcement learning, and rule-based safety checks. It evaluates on synthetic data and on TCGA ovarian cancer data: 587 patients and 2,552 treatment events.

00:21:20 The authors report better predicted benefit than baselines and 100 percent compliance with their coded constraints in experiments, while still asking for clinician review on uncertain cases. That doesn't grant permission to deploy. It is evidence that medical AI is moving toward systems that know when to stop and ask.

00:21:40 A physical-AI evaluation paper from XPENG Robotics argues that embodied systems need a shared evaluation runtime across foundation-model reasoning, visuomotor policy execution, and whole-body control. Their claim is that separate harnesses lose the shared trace needed to diagnose a regression that begins in one layer and appears in another.

00:22:01 The paper's phrasing is technical, but the consequence is plain enough: if a humanoid robot fails, you need to know whether the language planner, the visual policy, the controller, the simulator, or the resource layer caused it. A benchmark score without a trace isn't enough when the machine has mass.

00:22:20 Honda researchers published a crash-safety workflow for pedestrian protection that uses a surrogate trained on 1,800 computer-aided engineering crash simulations. They report an average R squared of 0.87 and say the workflow produced 35 safety-compliant alternatives in seconds, where conventional iteration would take weeks.

00:22:41 Then they add the sentence that matters: selected designs still require high-fidelity validation. Good. That is the contract. The model coordinates exploration; the physics still gets a vote. So after a day of leaders arguing about who gets access to frontier intelligence, the physical-world items bring the conversation back down to evidence.

00:23:03 Materials need experiments. Medicine needs clinical review and data governance. Robots need trace continuity. Crash safety needs high-fidelity confirmation. The next evidence I need is a written access process for frontier models and proof that systems moving into labs, hospitals, factories, and roads can preserve the record when something breaks.

00:23:24 Jonas