Sam Altman used to think the arrival of powerful AI would be an event. In a clip played on Moonshots with Peter Diamandis, the OpenAI chief executive described a specific expectation that did not come true: when GPT-4 arrived in 2023, he thought there would be "much more disruption in software business being up for grabs right away than turned out to be."
His explanation was not that the technology stalled. It was that customers did not move. "The economy just has so much inertia," he said. "People keep doing the same things they're doing. They keep buying from the same company. They keep sort of wanting to use their tools in the same way." He called that "actually a positive in many ways," said it would make the transition "go smoother and slower," and added, "I'm grateful for it." His conclusion: "we've all been too ambitious on timelines, even with this incredible technology."
Altman made those remarks in an interview with David Senra, where he also argued that continued technical progress can coexist with slower economic absorption, and described expanding computing capacity as a coordination problem spanning chips, fabrication, electricity, finance and logistics.
Diamandis framed the clip as technology hitting "the reality of society and humans" — the singularity as a process rather than a single moment. Then his panel spent the next several minutes disagreeing about which part of that reality is responsible.
Salim Ismail: the slow layer is where the stress is
Salim Ismail described a collision between two bottlenecks: the technical one and the one made of coordination, incentives and regulation. Frontier labs, he argued, make the mistake of "confusing technical possibility with institutional deployment" — two very different layers.
He reached for Stewart Brand's idea of pace layers, in which the parts of a civilization move at different speeds, like swift currents at the ocean's surface over slower water far below. Government and regulation, in his telling, sit in the slow layer, with good and bad effects. His example was personal: he had just spent two hours getting to the airport. Passenger drones, he said, have been technologically ready for a decade and are waiting on infrastructure and regulation; the trip could have taken ten minutes.
Brand's essay, published in January 2018, sets out six layers — fashion, commerce, infrastructure, governance, culture and nature. The fast layers experiment and absorb shocks; the slow ones hold memory and restrain destabilizing change. In his forest illustration, needles change from year to year, tree crowns and stands over longer spans, forests and biomes over centuries and millennia. For Brand the different speeds are part of what makes a system resilient, rather than simply obstacles to progress. Salim's interest is in what the slow layer is costing.
His conclusion redefined the word everyone else was using. The singularity, he said, is not the moment machines become infinitely capable; it is the moment institutions can no longer adapt to the rate of capability at all. "That is the breaking point. We're kind of there now." The remedy he wants is what he said Alex calls co-scaling: growing organizations and institutions fast enough to keep pace.
Alex: the villain is the abstraction stack
Alex agreed with Altman at the surface, and repeated his own view that treating the singularity as a step function is nonsensical — it is an interval of time, and we are in the middle of it. One layer down, he parted company with the diagnosis.
"I think the villain, if there is one in this story, is actually abstraction layers," he said. His example was the electric car. Swap an internal combustion engine for an electric one and, under the hood, the technology is a step change and completely unrecognizable. Go up one layer and it is still a car, with a recognizable steering wheel and recognizable wheels. The transformation is real and invisible, because the stack above it stayed the same.
On Altman's tone, Alex was skeptical of the gratitude. He read it as relief and complaint at once: "He has a way of sometimes like saying two things at once."
The theory comes with a prescription. If you want to go faster, Alex said, "pull an Elon and vertically integrate to erase the barriers between abstraction layers." Applied to OpenAI, that means reaching above the model layer or below it. He said the company had publicly abandoned its original Stargate strategy of owning its own data centers in favor of leasing, and argued that owning as much of the stack as possible — down to its own chips and the facilities that run them — is what would let it move quickly. He sees the labs beginning to do exactly that.
Emad Mostaque: the models only just got good
Emad Mostaque accepted the abstraction-layer point, then offered a third explanation that has nothing to do with institutions. Diffusion was slow, he suggested, because there was not much worth diffusing.
"The models weren't good enough until a few months ago," he said. The code they wrote a year ago was, relatively speaking, garbage; then it became okay; now he does not look at the code anymore. For mathematics, he named o3, roughly a year earlier, as the first model he could use at all, and the latest GPT-5-series release as the first really good math model. Wrapping competent intelligence in familiar interfaces — a chat window, a messaging thread — is something he dated to perhaps a month or two. His test of the earlier era was blunt: he would not use GPT-4. "Can you imagine using GPT4 in a code base?"
He also noted the awkwardness of the moment: a Time article he said he had not read, in which Altman put AGI at the end of this year, running alongside Altman's surprise at how slowly the technology has spread.
Oversupply, and the $30 trillion question
Diamandis offered his own metaphor, an impedance mismatch: increasingly powerful tools meeting governments, companies or individuals that cannot take advantage of them. His view of the options was stark — hand the AI an objective function and turn the job over to it, or keep a human at the interface, where, he said, the human breaks quickly.
Alex converted the same idea into economics. OpenAI and Anthropic, he said, largely have an oversupply of intelligence. From participating in and watching the market over recent months, his impression is that not all of the market has demand for it, is ready to have that demand, or knows how to use it when the supply is there. The two curves from economics 101 are not crossing everywhere at once.
The panel then turned to where that leaves the labs. One panelist, who said he is chairman of a large public insurance company, expects frontier companies to reach up the stack and build or partner into fully verticalized finance, insurance and consulting businesses. His own company's attention has already shifted: a year ago it cared about auto insurance; now the new categories AI is generating — data centers, robots — look bigger than the legacy insurance industry. Rather than disrupting incumbents who get angry and vote against you, he suggested, the new economy can simply be built alongside the old one.
Alex called the direction of travel the $30 trillion question. Facing margin pressure on model releases, is it more natural for an American frontier lab to go upstack or downstack? More ergonomic to go downstack, he guessed: design their own chips, compete with Nvidia, operate their own data centers and energy — though he expects them to try all of it. Diamandis added that Anthropic has put its total addressable market at $30 trillion, and wondered aloud where the number came from, noting the coincidence that US GDP is about the same figure.