IMPULSE — May 4, 2026: Ten Billion Miles, Zero Percent China

00:00:04 Today's show covers a week in which a lot of institutions outside the AI industry shifted what AI is doing to the world: the federal government, the largest healthcare buyer on the planet, the largest electric vehicle company, the largest AI chip maker, an emergency room in Boston, a research lab in Yorktown Heights, and one of the more sober frontier-lab policy voices in San Francisco.

00:00:26 The questions on the table are whether American patients can keep their citizenship status off an ad exchange, whether a self-driving company gets to define its own safety story, whether the world's largest semiconductor market is now closed to American silicon, and whether the next research breakthrough comes out of a graduate seminar or out of a model that wrote its own training run at three in the morning.

00:00:50 I'm Jonas Vale, and this is IMPULSE for May 4, 2026.

00:00:53 The first item is policy, and it signals more than it commits to. According to reporting picked up by Hacker News yesterday, the Office of Science and Technology Policy is in early conversations with at least three frontier labs about a voluntary pre-release evaluation regime.

00:01:10 The shape of the proposal, as best I can tell from the secondhand summaries, is this: before a lab ships a model that crosses some threshold of capability — the threshold itself unspecified — they would submit weights, evaluation harnesses, and red-team artifacts to a federally coordinated test bed, run by NIST in cooperation with the AI Safety Institute, with results held under nondisclosure for some window.

00:01:35 None of this is law, or even a written executive action. It is, in the language of people I have heard from on the Hill, a temperature check. What changed is that the temperature is being taken at all. For most of 2024 and 2025, the federal posture toward frontier model release was hands-off in the technical sense and aggressive in the antitrust sense — the FTC asking questions about training data and partnerships, but no one in the executive branch suggesting that a model should sit in a queue before shipping.

00:02:06 That is what is being floated now. The Hacker News thread on this is unusually substantive. The top comment, from someone who identifies as a former FDA reviewer, draws the comparison to the 510(k) medical device pathway and notes, with the careful understatement of someone who has watched a regulatory process up close, that the 510(k) program took roughly fifteen years to find its working equilibrium and is still litigated annually.

00:02:33 The reply underneath, from a current frontier-lab safety researcher posting under their own name, is shorter. It says: we already do this internally; the question is whether the federal version adds signal or just adds latency. Both of those people are right. What I want to know — and I do not yet have an answer to this — is what counts as the trigger.

00:02:54 Is it parameter count? Compute budget? A capability score on a specific eval? Each one is gameable, and each one shifts the incentive structure of model design in a different direction. If it is compute, you get more efficient training runs and the same capability.

00:03:10 If it is a capability score, you get models deliberately tuned to underperform the trigger eval. The history of regulated industries is the history of compliance teams learning the test. There is no reason to think frontier AI will be different. The thing to watch over the next quarter is whether any of the three labs reportedly in conversation — the reporting names two of them, Anthropic and OpenAI, but is vague on the third — agrees to a public commitment.

00:03:38 A public commitment is what turns a temperature check into a precedent. Without one, this is a story about a meeting. With one, it is a story about a regime.

00:03:47 The second item is the kind of story that does not get the attention it deserves because it is technical and slow and involves the words enrollment broker. A research group at the Markup, working with a coalition of state insurance commissioners, published findings yesterday that the federal healthcare exchange — healthcare dot gov — and at least eleven state-run marketplaces have been transmitting personally identifiable enrollment data, including citizenship status, race, ethnicity, and in some cases income tier, to third-party ad tech vendors via tracking pixels embedded in the application flow.

00:04:23 This is a violation of the HIPAA Privacy Rule on its face. Depending on the state, it is also a violation of state-level marketplace contracts that explicitly prohibit data sharing with non-business associates. The companies named in the report include the usual ad tech suspects — Meta, Google, and a handful of programmatic exchanges that route through demand-side platforms most listeners would not recognize.

00:04:47 The top Hacker News comment on this is from someone who works in healthcare compliance and says, and I quote, this is not new. The pixel tracking issue has been documented since 2022. The reason it is news today is that the federal exchange is now in the dataset, which means CMS has direct liability under the Office for Civil Rights enforcement framework.

00:05:08 The detail you do not hear in most of the coverage is that the AI angle here is not the data collection itself. It is what happens downstream. Ad networks build lookalike models. Lookalike models train on whatever attributes they are fed. If you fed an ad network the citizenship status of every person who applied for federal health coverage between 2023 and 2025, that ad network can now infer citizenship status on every other user in its graph with non-trivial accuracy.

00:05:36 Even if the original data is deleted under a consent decree, the inference capability persists in the model weights. This is the part the privacy litigation framework is not built for. You can sue for the data transmission and extract a fine. You cannot un-train a lookalike model.

00:05:53 The Markup report does not make this point directly — the authors are careful to stay within what the documents support — but every machine learning person I have spoken with in the last twenty-four hours has said some version of the same thing: the pixel was the symptom, and the trained inference is the disease.

00:06:11 What I expect is a settlement that includes some form of model retraining covenant. There is precedent for this from the FTC's case against Cambridge Analytica's algorithmic descendants, and from a 2023 settlement involving a facial recognition vendor that was required to delete not just the dataset but any derived model artifacts.

00:06:31 That precedent is fragile. It has not been tested in federal court at the scale of an ad network. If this case is the one that tests it, the outcome will reshape what privacy law actually means in a model-training era. If the case settles for a fine and a promise, it will reshape nothing.

00:06:48 We should know inside of six months.

00:06:50 Tesla announced yesterday that the cumulative mileage across all vehicles operating in Full Self-Driving Supervised mode has crossed ten billion miles. The announcement was a single graphic posted to the company's account, with no accompanying safety report, no breakdown by software version, no disengagement statistics, and no comparison to human-driven baselines.

00:07:12 The CEO retweeted it with the word inevitable. Ten billion miles is, by an order of magnitude, the largest dataset of supervised partial autonomy ever assembled. Waymo's fully autonomous fleet, by their last public number, has accumulated something on the order of seventy million driverless miles.

00:07:30 Cruise, before its operating suspension, was in the low tens of millions. Mobileye's supervised system, deployed across a number of legacy automakers, is in the same ballpark as Tesla but is not unified across software versions in the same way. So the headline is real.

00:07:46 Tesla has more miles of someone-pressing-the-pedal-while-the-car-steers than anyone else on the planet, by a wide margin. It is not a safety claim. It is a number a company prints because it is allowed to. The federal regulatory framework for partial autonomy in the United States, governed by NHTSA's Standing General Order on crash reporting, requires manufacturers to report crashes involving Level 2 and above driver assistance systems within set timeframes.

00:08:14 It does not require them to publish miles driven, disengagement rates, the population of drivers, the geographic distribution of usage, or the version of software in operation at the time of any given incident. NHTSA collects some of this and publishes very little of it.

00:08:30 The result is that the company can pick its own metric. I do not think Tesla's FSD system is unsafe. I do not have the data to think anything definitive about it. That is the point. The number that matters — the rate at which FSD-engaged miles produce serious injuries or fatalities, normalized for road type and driver demographic and weather, compared to the human baseline — is not a number anyone outside the company has.

00:08:55 It is, at this point, a number that NHTSA itself does not publish in the form that would let an outside researcher answer the safety question. The closest public proxy is the company's quarterly Vehicle Safety Report, which is not audited and uses an internal definition of crash that the company controls.

00:09:13 This is what the deployment of physical AI looks like in the United States today. A company crosses a milestone, defines the milestone, and prints the graphic; the regulator collects some data and publishes very little of it; the press writes the headline the company wrote.

00:09:30 The population of people whose roads the cars are driving on — pedestrians, cyclists, drivers of other cars — has no standing to ask for the breakdown. That last group is the one I keep coming back to. The driver who buys a Tesla has consented to FSD by buying it.

00:09:46 The pedestrian crossing the street has consented to nothing. There is no framework in American transportation law that recognizes the second consent as a thing that needs to exist. That is the gap the ten billion miles is sitting on top of.

00:10:00 Jensen Huang said something striking in an interview yesterday, and the clip is being passed around in policy circles for reasons that go beyond the chip business. Speaking on a CNBC segment about the next-generation export control framework, the Nvidia CEO said his company's market share in China for advanced AI accelerators is, quote, zero percent, and that the company is, quote, no longer competing in that market in any meaningful sense.

00:10:27 The framing matters. He did not say share has declined or that share is small. He said zero. It is worth being precise about what that claim covers. Nvidia still sells products into China — gaming cards, professional workstation GPUs, and the export-compliant H20 derivative that the company spent most of 2024 and 2025 trying to get cleared for shipment.

00:10:48 What the CEO is saying is that in the segment that matters strategically, advanced training accelerators for frontier model work, Nvidia has effectively withdrawn. The buyers in China are now Huawei, with the Ascend 910C, and a small number of domestic startups working on more specialized inference silicon.

00:11:07 The Financial Times reporting in March put Huawei's market share in Chinese AI training silicon north of seventy percent. The CEO of Nvidia is saying the remaining thirty is not them. This is a calculated public statement, and I think the calculation has two audiences.

00:11:23 The first is the US government, which is in the middle of revising the export control framework yet again, with the Bureau of Industry and Security signaling new rules around end-use verification and offshore compute access. By saying we have already lost this market, Nvidia is making the argument that further restrictions are now costless to American industry — there is nothing left to give up — and so the case for tightening is purely strategic, not economic.

00:11:51 The second audience is the investor base, which has spent a year worrying about whether the company's revenue is exposed to a Chinese policy reversal. Saying it is at zero is saying the downside is bounded. The question underneath the framing is what the Chinese frontier labs are now doing with the silicon they have.

00:12:10 There is a thread on Chinese AI infrastructure that I have been following for several months, and the rough picture is this. The Ascend 910C is a real chip. It is not as performant per watt as a Blackwell, but it is good enough for training models in the GPT-4 capability class, and Huawei has been able to scale production faster than US analysts predicted in 2024.

00:12:32 The Chinese open-weight ecosystem — DeepSeek, Qwen, GLM — has been training on this domestic silicon for a year. The models are competitive on benchmarks, and the cost structure is different from the American labs because the silicon is subsidized. The next thing to watch is the next major Chinese frontier release, and the disclosure of what hardware it was trained on.

00:12:54 If a top-tier model ships from a Chinese lab in the next two quarters, trained entirely on Ascend, that is the moment the geopolitical story shifts from chip access to chip independence. The American export control regime was built on the assumption that withholding silicon would slow Chinese frontier progress by years.

00:13:14 If the slowdown turns out to be measured in months, the policy framework is not just outdated; it is counterproductive. That is the conversation that will be happening in Washington over the summer.

00:13:26 A study out of Harvard Medical School and Mass General, published in the New England Journal of Medicine and covered by TechCrunch yesterday, ran a frontier reasoning model — the paper does not name the model commercially, but the methods section makes clear it was a recent OpenAI o-series checkpoint — against a panel of board-certified emergency medicine physicians on a battery of seventy-nine de-identified case presentations from the hospital's actual triage records.

00:13:53 The model's diagnostic accuracy on the primary diagnosis was, by the paper's measure, seventy-six percent. The physician panel's median accuracy was sixty-three percent. The gap is statistically significant. The sample size is small. The lead author, Dr. Adam Rodman, gave a quote to TechCrunch that I want to read in full, because it is the kind of statement from a clinician researcher that does not usually make it into AI coverage.

00:14:17 He said, quote, the result is striking, and we should be careful about what it means. The model is doing something different from what the physicians are doing. The physicians are working with incomplete information under time pressure with a queue of other patients behind them.

00:14:33 The model is working with a clean text vignette and unlimited reasoning time. We have not shown that the model is a better diagnostician in the ER. We have shown that on a structured retrospective task, it produces a more accurate first guess. Those are not the same thing.

00:14:49 End quote. Rodman is the right kind of cautious, and the study itself is more careful than the headlines about it. The thing that interests me is not the seventy-six versus sixty-three. It is the failure mode comparison. The paper has a table — table four, if you go look — that breaks down the cases where the model and the physicians disagreed.

00:15:09 In the cases where the model was right and the physicians were wrong, the dominant pattern was that the physicians anchored on the chief complaint and the model considered a wider differential. In the cases where the physicians were right and the model was wrong, the dominant pattern was that the physicians integrated a piece of bedside context — a patient's affect, a family member's offhand comment — that did not appear in the structured vignette and could not have been available to the model.

00:15:37 That asymmetry is what the deployment conversation should be about. What does this mean for medicine? In the near term, very little. The FDA's Software as a Medical Device framework does not have a clearance pathway for a general-purpose reasoning model used for diagnostic support, and the major EHR vendors are nowhere close to integrating one in a way that would survive a malpractice deposition.

00:16:00 In the medium term, I would expect to see narrow-task clinical decision support tools that wrap a frontier model in a guardrail layer and submit for clearance under existing pathways. Some of these will be useful and some will be embarrassing. The question for hospital systems is not whether to adopt — they will, eventually, all adopt — but whether the adoption happens through a vendor procurement process that includes safety review, or whether it happens through individual physicians pasting case notes into a consumer chatbot in the hallway.

00:16:31 That second pattern is already happening, and it is the one the hospital legal teams are losing sleep over.

00:16:37 IBM Research published a paper and an open-weight checkpoint last week, and the Reddit MachineLearning thread on it has been climbing all weekend. The model is called MAMMAL — the acronym is for Multimodal Aligned Molecular Architecture for Life sciences — and the claim is that it is a single transformer trained jointly on protein sequences, DNA sequences, small molecule SMILES strings, and a structured representation of histopathology image patches.

00:17:04 The training corpus is in the range of two trillion tokens across all modalities. The model is, by the paper's benchmarks, competitive with or ahead of the modality-specific specialist models on a number of standard tasks — protein function prediction, drug-target binding affinity, variant pathogenicity classification — while being a single set of weights.

00:17:25 The institutional fact about MAMMAL is not the architecture. It is the release. IBM has chosen to release the weights under a research license that permits academic and non-commercial use, including commercial drug discovery work by qualifying biotech startups under a fee structure the paper describes as nominal.

00:17:43 This is a different posture from the one the major American AI labs have taken on biological models. Anthropic and OpenAI do not release biological capabilities at all. Google DeepMind released AlphaFold 3 weights only after sustained pressure from the academic community and only under a restrictive non-commercial license.

00:18:02 IBM is shipping a more capable cross-modal model on materially better terms. Why does this matter outside the molecular biology community? Because drug discovery is one of the few areas where AI capability has a direct and measurable line to human welfare, and because the access regime around the foundation models is going to determine who gets to build on top.

00:18:24 If the best multimodal biological model in the world is gated behind a frontier lab's commercial agreement, the population of researchers who can use it is small and concentrated in well-resourced institutions. If it is open under a research license, the population is large and includes academic groups in countries that cannot afford a frontier lab's pricing.

00:18:45 The pace of biomedical research is set, more than people in this industry like to admit, by who has access to the tools. I am not in a position to evaluate the science of MAMMAL on its merits. The paper is dense, the benchmarks it reports are the standard ones, and the standard ones in this domain have known limitations.

00:19:04 What I can say is that the release strategy is a deliberate signal from IBM that the company is going to compete in foundation models on access rather than on raw capability. That is a coherent strategy and a different one from the strategy of every other lab in the top tier.

00:19:20 Whether it produces the science that justifies the position is the empirical question of the next two years.

00:19:27 Jack Clark, who runs policy at Anthropic and writes one of the more careful weekly newsletters in the field, gave a talk at a closed-door conference last week, and a Reddit thread surfaced a partial transcript yesterday. The line being passed around is his prediction that by 2027 or 2028, frontier AI systems will be capable of, quote, conducting meaningful AI research themselves — proposing experiments, running them, interpreting the results, and incorporating the findings into a next training run, end quote.

00:19:54 The qualifier he added, which is being cut from most of the social media versions, is that this is conditional on continued scaling of inference compute and on no major regulatory or supply-side disruption. Be specific about what the claim is and is not. It is not the claim that AI systems will be doing all AI research.

00:20:12 It is the claim that they will be capable of conducting some research with meaningful results, in the way that a strong graduate student is capable of conducting some research with meaningful results. The window — 2027 to 2028 — is roughly two years out. The mechanism is not that current architectures suddenly become better at it.

00:20:30 It is that the inference compute available per query keeps growing, and that the harness around the model — the tool use, the experiment management, the result interpretation — keeps maturing. Clark is making a forecast that combines a hardware trend with a software trend.

00:20:44 If this prediction is right, the implications are not primarily technical. They are institutional. The current structure of frontier AI research is built around a small number of labs that hire a small number of very expensive researchers and give them access to very expensive compute.

00:21:00 If the marginal cost of running an experiment drops by an order of magnitude because the experiment is now being designed and executed by a model rather than a person, the bottleneck shifts. What becomes scarce is not researcher time. It is compute, and it is the judgment about which experiments are worth running.

00:21:17 Both of those are concentrated in the same places they are concentrated now, but the leverage per dollar goes up substantially. The other implication, the one Clark himself is careful about in the longer version of his remarks, is on the safety side. If a model is proposing experiments that go into the next training run, the question of what the model is selecting for becomes the question that determines what the next model is.

00:21:40 The current alignment toolkit assumes a human researcher in the loop deciding what to train on. If the human researcher is a reviewer rather than an originator, the failure modes shift. He does not have a complete answer to this. Nobody does. The marker to watch for is the first publishable result that came out of a model-conducted experimental loop, with a paper trail showing what the model proposed, what it ran, and how the results were validated.

00:22:05 That paper, if and when it appears, will be the first piece of evidence on whether the 2027 timeline is a forecast or a marketing line. I am skeptical of compressed timelines as a default. I am also aware that the people running the labs have access to internal capability evidence that the rest of us do not.

00:22:21 Clark has been more measured than most in public. When a measured voice gives a specific date, I'll mark it and check back. That is what I am watching tomorrow. Jonas.