Week One in Oakland, Many Jobs Go Away, and the Bubble That Wasn't

00:00:04 From IMPULSE, this is Jonas Vale. It is Friday, May first, twenty twenty-six. Today: a federal courtroom in Oakland, where Elon Musk admits under oath that his AI company trains on OpenAI's outputs. Sam Altman posts two lines about jobs going away. The Atlantic explains how we got from 'AI is a bubble' to 'not enough data centers' in six months.

00:00:27 And a Nature paper out of Hong Kong and Boston describes a way to deploy a medical foundation model in a clinic without paying for frontier-lab inference on every query.

00:00:38 A federal courthouse in Oakland, California, this week — packed with lawyers carrying boxes of exhibits, a row of OpenAI employees, and protesters outside holding signs telling people to quit ChatGPT or boycott Tesla or both. That is where Elon Musk took the stand in Musk versus Altman, the trial that could unwind OpenAI's for-profit restructuring and remove Sam Altman and Greg Brockman from the company.

00:01:04 Musk's framing was that he was duped. Quote: 'I gave them thirty-eight million dollars of essentially free funding, which they then used to create what would become an eight-hundred-billion-dollar company.' He is asking the court to remove Altman and Brockman from their roles and to reverse the move from nonprofit to capped-profit to whatever the structure is currently called.

00:01:27 The remedy lands just as OpenAI is reportedly heading toward an IPO at a valuation approaching a trillion dollars. Meanwhile, his own xAI is expected to go public as part of SpaceX as early as June, with a target valuation of one-point-seven-five trillion. A few items from the week.

00:01:45 Judge Yvonne Gonzalez Rogers snapped at both sides when the lawyers started arguing about who is the truer steward of AI safety. Quote: 'This is not a trial on whether or not artificial intelligence has damaged humanity.' She also said, of Musk himself, 'I suspect there's plenty of people who don't want to put the future of humanity in Mr.

00:02:06 Musk's hands.' And then the part that drew gasps in the courtroom. Under cross-examination by William Savitt, who once represented Musk and Tesla, Musk admitted that xAI 'partly' distills OpenAI's models. Distillation, in this context, means using a larger model's outputs as a training signal for a smaller, cheaper one.

00:02:26 This is the same practice OpenAI publicly accused the Chinese AI company DeepSeek of in February, and the same one Anthropic blocked OpenAI from doing to Claude last August. Musk's defense was, quote, 'It is standard practice to use other AIs to validate your AI.' Validate is a generous word for what distillation actually does.

00:02:46 What is being decided here is corporate control of the leading US AI lab. If the court grants any meaningful piece of what Musk is asking — remove the founders, unwind the structure — that is a pre-IPO governance event of the kind we usually see when a regulator acts, not a competitor.

00:03:04 Stuart Russell, the Berkeley computer scientist, is set to testify on AI safety next week. Brockman, who has been taking notes through Musk's testimony, is also expected to take the stand. I want to hear what Brockman says, under oath, about how the for-profit conversion was actually presented to the original donors.

00:03:24 From the courtroom to a tweet. Sam Altman, on Twitter, late Friday Eastern, in two lines. Quote: 'many current jobs will go away. i think we will find a lot of new ones, though they may look very different.' That is the whole post. I do not want to over-read a tweet — it is a tweet — but it is a softer phrasing than the version of Altman who, a couple of years ago, used to call out his peers for downplaying the labor risk.

00:03:51 The top reply, from Tyler Johnston, made exactly that point: 'I honestly miss the Sam Altman that used to call out his peers for downplaying this risk.' The next reply down was an unemployed user from India asking, plainly, which jobs. There were a lot of those.

00:04:07 A former game developer wrote, 'I bust ass to learn how to make AI and sharpen the blade, and I'm pretty sure none of it will matter.' A reply from a manufacturing-focused account wrote that the question is not whether new jobs come, it is whether they come fast enough.

00:04:24 And one reply, from an account called MachineSovereign, asked the institutional question. Quote: 'The harder question is whether jobs remain the central mechanism for distributing income, status, identity, bargaining power, household formation, and a claim on the future.' That is not a Twitter argument.

00:04:44 What holds the social structure together if waged work shrinks faster than the new categories grow — that is a labor-economics and welfare-state question. There is no US plan for it. There is no UK plan, German plan, or Japanese plan. There is a sentence on Twitter from the CEO of the company most aggressively rolling that change out, saying it will probably be fine.

00:05:07 If labor is the demand-side question, the supply side is compute. Six months ago, every other piece in the financial press said 'AI bubble.' Now the same press is writing about a data-center shortage. The Atlantic ran a piece this weekend laying out why. Ethan Mollick, the Wharton professor who is quoted in it, posted his own one-line summary.

00:05:29 Quote: 'It lays out the reasons why we whipsawed from AI is a bubble to there are not enough data centers in less than six months. Spoiler: it's agents.' The math is plain. A chatbot is one inference per interaction. An agent doing a real task — closing a ticket, drafting a contract, processing a refund, writing and testing a function — runs dozens of inference calls per task, on its own schedule.

00:05:54 One reply on Mollick's thread put a number on it: fifty-plus calls per task, autonomous, twenty-four seven. Another reply, less polite, said: 'one tiny loop can quietly become two hundred calls because it kept rereading the same repo.' Both are right. So the bubble framing and the shortage framing are not actually contradictory.

00:06:15 One reply made the historical comparison I would make: the eighteen-forties railway bubble. The stocks crashed. The tracks remained. A transportation revolution happened on top of the tracks. What is plausible here is that several public-market valuations are wrong by an order of magnitude in either direction, and the underlying compute commitment continues to be real because enterprise procurement has signed the checks.

00:06:42 Hyperscaler capital expenditure this quarter is, again, very large. The question I would ask is not 'is it a bubble.' It is: who owns the data centers when the equity correction comes, and which jurisdictions get the next wave of buildout. That is a power question, not a price question.

00:07:00 Power and price both run through compute, and the last item is about a way to use less of it. Out of the Hong Kong University of Science and Technology, Harvard Medical School, Weill Cornell Medicine, and a half-dozen Chinese hospitals, a paper in Nature Biomedical Engineering this week describes what they call Generalist-Specialist Collaboration.

00:07:22 The architecture is straightforward. A large open-source medical foundation model called MedDr, paired with a suite of small, narrow specialist models — one for retinal images, one for skin lesions, one for chest X-rays, one for pathology slides. When a case comes in, the specialists give the generalist their diagnostic predictions and a handful of visually similar prior cases as context.

00:07:46 Then the generalist makes the call. They evaluated it on thirty-two datasets across radiology, dermatology, pathology, ophthalmology, and endoscopy. They report MedDr-plus-specialists outperforming both the standalone generalist and the standalone specialists on diagnosis and report generation.

00:08:04 Most of what gets written about AI in medicine is one of two things: a hospital signing a contract with a frontier lab, or a benchmark headline that does not survive a real clinical workflow. This is something different. It is a paper saying you can deploy a medical foundation model at clinical scale without burning frontier-lab compute on every query, by using cheap, narrow specialists as a context layer.

00:08:29 The weights and code are public on GitHub. That is the kind of architecture that ends up running in a hospital in Shenzhen or Hyderabad before it runs in one in Boston, because the local compute budget actually allows it. I have not seen independent replication yet.

00:08:45 I would want a clinician, not a paper, to tell me whether the report-generation outputs hold up in chart review. But the deployment shape is what I'll be watching.

00:08:55 Four stories from the day. A courtroom, a tweet, an Atlantic piece, and a Nature paper. The thread, lightly: who controls the leading US lab, who carries the labor risk, who pays for the compute, and who can deploy AI in a clinic without a hyperscaler bill. Next week I am watching Greg Brockman's testimony in Oakland.

00:09:11 Jonas.