Infrastructure Hiring Is Delaying AI’s Broad Jobs Reckoning

John Coogan treated the claimed Navier–Stokes breakthrough as a test of what people mean when they say AI progress “matters.” The equations govern familiar physical problems—airflow over wings, weather, and water moving through pipes—but, as Tyler put it, engineers and physicists already solve them numerically in particular cases. What remains unresolved, in the framing they discussed, is a general mathematical proof.

That distinction is central to Coogan’s skepticism about practical claims. Better weather prediction, more efficient aircraft, longer ranges, lower fares, and lower emissions are all intuitively appealing consequences to attach to Navier–Stokes. But Coogan said that is not what a proof would deliver. Physicists, he said, had chimed in with the view that proving the equations would not measurably advance engineering. It is a “cool demo,” not a direct route to better wing design or less turbulence.

An on-screen TBPN graphic described the claim in concrete terms: “OpenAI internal model produces 165-page Navier-Stokes proof.”

For Coogan, the significance of such a result is primarily evidentiary about model capability: a hard benchmark that makes the pace of AI progress legible to technically minded observers. Last year’s International Math Olympiad results played a similar role. Coogan said Google DeepMind and OpenAI each reached 35 out of 42, while no AI model solved question six. He also recalled that Scott Wu had predicted on TBPN early in 2025 that AI would achieve an IMO gold-medal-level result that year.

The race has made evaluation and credit contentious. Google’s result, Coogan said, was graded using the official rubric for that year, while OpenAI had former IMO gold medalists validate its work. His analogy was that one team had run the fastest 100 meters “in the parking lot” while the actual Olympics took place in the stadium. In Coogan’s account, similar disputes now surrounded the claimed Navier–Stokes work: mathematicians and researchers were arguing over whether data had been improperly used, whether work had been copied, whether the approaches were actually distinct, and who deserved credit.

Tyler nevertheless argued that usefulness is too narrow a standard. A Millennium Prize problem belongs, he said, more to the category of a beautiful painting than a practical engineering intervention: pure mathematics at “the peak of the mountain.” Coogan did not reject that view, but separated mathematical achievement from the public experience of AI. Most people, he suggested, will find model specifications, compute expenditures, and formal proofs less compelling than an output they can immediately see and use.

Coogan expects the public narrative to return to biology and medicine after major open math problems are resolved. Yet he also cautioned that “curing cancer” will not provide a clean public-relations breakthrough for AI labs. Even a 20% improvement against one particular cancer, Jordi Hays argued, could register as another incremental clinical advance rather than a sweeping technological victory. Drug development, testing, and observing outcomes impose a slower cycle than spending heavy inference budgets on an open mathematical question over a weekend.

John Coogan repeatedly returned to a category of evidence more immediate than a formal benchmark: systems doing recognizable work in the world.

He saw 3D generation as more visceral than abstract model results. Image and video generation were already familiar; generating a navigable 3D model from a few images felt different because a person can inspect it from every angle. The most impressive examples, he argued, are not polished objects that already exist online—a train with abundant existing 3D assets, for example—but a specific house, car, or other object with no ready-made model to borrow.

That concreteness extends beyond graphics. Coogan described Astra as compelling for computer use: not merely answering a question, but traversing settings menus, command-line tools, and control panels to make a requested configuration change. Hays’s mundane example was a parking ticket that has not yet appeared in the relevant portal. Instead of remembering to check repeatedly, a user could tell an agent to wait until the ticket becomes available and then pay it.

That is a modest form of “superintelligence,” as Hays jokingly called it, but it illustrates a more consequential shift than a one-off answer: an agent can hold a task open, monitor for a prerequisite, and act when it arrives. The value comes from delegated persistence. A system does not need to be independently brilliant to remove the administrative burden of repeatedly checking a portal.

A timelapse showed a robotic arm painting the Golden Gate Bridge in four attempts, explicitly labeled “Attempt 01” through “Attempt 04.” The image improved over the sequence. Coogan thought the first attempt was rough; he and Hays considered the later attempts impressive. His point was not that the system had become a master painter. The task was immediately intelligible: a machine perceives, controls a brush, makes an attempt, and gets better.

Coogan also raised the possibility that generated 3D models could contribute to a 3D-printing boom. He described another demonstration that went from a prompt to a LEGO-kit-like design in one pass. If that workflow produces usable designs, models become inputs to physical fabrication rather than solely digital images. A separate anatomy site depicted a male body broken into 2,234 modeled pieces. Coogan questioned how much had been individually modeled versus taken from an existing model and animated, but said it still worked as a compact demonstration of why the technology might get someone’s “wheels turning.”

A separate Astra demonstration turned the usual CAPTCHA into a boss battle: visitors had to defeat an enemy to submit a contact form. The joke rests on a real tension. If agents can defeat conventional friction mechanisms, service providers may need stranger—or more capable—ways to screen incoming requests.

Hays connected that problem to Instinct, a consumer personal-assistant system that connects to email and text messages. In one example discussed, a user was reportedly banned from Resy after an agent relentlessly pursued restaurant reservations. In another, a user who appeared briefly on a US Open fan camera asked Instinct to find the footage. Hays said the system contacted hundreds of media contacts and followed up aggressively; it ultimately got the video. Coogan called the outreach “exceptionally annoying,” even as he acknowledged that it succeeded and produced a mutually beneficial outcome.

The problem is therefore not simply whether an agent completes the task. It is who bears the cost of its persistence. Coogan’s proposed answer was recipient-side agent intermediation: a system on the receiving end could recognize that a request resembles spam but contains a legitimate action, process it without forcing a person to endure the barrage, and let the useful request through. In that account, agentic systems do not reduce demand on organizations so much as create a need for systems that negotiate, filter, and prioritize on their behalf.

Jordi Hays and Coogan saw a familiar problem in the new run of technology-industry films: a work intended as a warning can make its subjects more compelling.

That was the question they attached to Artificial, whose trailer paired a promise of a machine that would solve the world’s problems with warnings that countries and industries would collapse. Coogan thought the visuals were striking, but wondered whether the film would seem too soon, unnecessary, divisive, or oddly inspiring to entrepreneurs.

Hays framed the concern through The Social Network. A film can present Mark Zuckerberg as troubled or socially alienated and still leave viewers thinking, “Wow, I wanna be like Mark Zuckerberg.” Neither host claimed to know whether Artificial would produce that reaction. Their point was that audience reception can outrun a work’s apparent moral intention.

They expected the Elizabeth Holmes documentary You Can See Everything to have a broader cultural base than a narrowly tech-industry audience. Its trailer showed Holmes insisting she had no reason to deceive, followed by Nathan Fielder responding, “The acting is so good.” Coogan said the Theranos story had reached well beyond specialist audiences through broad news attention and John Carreyrou’s book. He also stressed how little was known about the documentary itself beyond footage filmed after Holmes’s sentencing and before she reported to prison, plus recorded visits by Fielder to the facility.

They were more doubtful about The Social Reckoning. Coogan read its teaser as centering on Frances Haugen, Facebook’s internal documents, congressional testimony, misinformation, and claims about teenage girls’ mental health. He wondered, explicitly as speculation, whether its timing was meant to precede addiction-related trials. Hays predicted it would flop because the story was no longer top of mind; Coogan agreed that viewers may be more interested in current AI talent wars than a return to platform-era controversy.

John Coogan closed on a counterweight to claims of imminent AI-driven unemployment. The headline he cited from The Economist was direct: “The jobs apocalypse is postponed. An AI jobs boom is here.” Its case, as Coogan summarized it, was cautiously positive. AI may eventually make many people unemployable, but there is no sign of it yet.

He cited reported August figures from the Bureau of Labor Statistics: 162,000 jobs added, above expectations, and a 4.1% unemployment rate. He emphasized that young workers, often framed as automation’s first likely casualties, were holding up comparatively well; the unemployment gap between 20-to-24-year-olds and the population overall was said to be near a multidecade low.

The aggregate picture does not deny disruption. Professional and business-services hiring was running 10% below its 2015–19 average, according to Coogan’s summary. Microsoft and Meta were reducing headcount as they reorganized around AI. Block and Intuit were described as replacing some people with bots. Challenger, Gray & Christmas data put announced AI-related job cuts at roughly 16,000 a month so far that year.

But Coogan stressed scale. The American labor market typically churns through 1.7 million jobs lost and 1.8 million jobs added in a month, he said. On that basis, AI-related cuts make up less than 1% of ordinary monthly job losses and firings.

The charts separated AI-related employment into two categories. One covered “hard-hat labour,” including electrical contractors and HVAC and plumbing workers; the other covered white-collar STEM roles, including software developers and mathematics and data-science workers. That supported Coogan’s account of where job creation is occurring: data-center and power-generation investment creates demand for construction and infrastructure labor, while AI companies and incumbents expand technical hiring.

Coogan cited economist Kevin Bryan’s sarcastic reaction to the apparent surprise: a productivity-enhancing, investment-supporting technology should, absent other information, be expected to help workers initially. The reasoning is that new technology drives investment, and firms hire to pursue the opportunities it opens.

That is not a declaration that displacement will never arrive. It is a narrower claim about the evidence available now: particular companies, back-office functions, and professional services are being affected, but the broad labor market has not produced the apocalypse many expected. As Coogan summarized the current balance, AI is cutting jobs in some places while infrastructure buildout, startup hiring, and new technical roles are creating more than it has eliminated.