Dave, an entrepreneur on the Moonshots with Peter Diamandis panel, says he keeps putting the same question to other founders, and that almost nobody has a good answer to it. Suppose he handed them 100,000 genius-level employees tomorrow, agents that would follow their exact marching orders. What would they set them to work on?
"It's a very hard problem because we've never had that opportunity before," he said. "We don't think about it a lot." Most people, in his description, are still using AI as a co-pilot — an assistant sitting beside one person doing one person's work — and conclude from that experience that they know what AI is. Very soon, he said, it will be 5,000 concurrent agents, then 10,000, then 100,000. Translating that into "I want a better humanity. What would I do?" is the part people cannot do.
The reason the question had teeth that morning was a result announced days earlier. OpenAI said it had used roughly 10,000 concurrent agents to produce a claimed proof about the Navier–Stokes equations — the mathematics of how fluids move, which the Clay Mathematics Institute illustrates with boat wakes, breezes and the turbulence behind an aircraft, and which sits on its list of seven Millennium Prize problems. The company reports about 88 hours to the result, another 17 hours to check it in the proof language Lean, and says it does not intend to claim the prize.
The easy part was saying what to do
What struck Dave about that run was not its difficulty but its clarity. Take the highest-level foundation model, deploy a couple thousand copies, or ten thousand, on different routes to the same target, and one of them comes back with a solution. By entrepreneurial standards, he said, that is "particularly easy." It is a hard problem, "but specifying the problem is really pretty damn easy."
Now try the other kind: solving cancer, he said, or better building construction, or better flight and travel plans for a colleague. Pointing the same machinery at those is much harder, and that is "the entrepreneurial journey that matters right now."
Diamandis recognized the difficulty from his own work. "This is what we say at XPRIZE," he said — the hardest part is defining a great challenge, a target to shoot for. Dave added the line the segment kept returning to: "those targets are not cooked."
His encouragement was that this is an unusually friendly problem to work on. Nobody is trying to stop you; the foundation model companies want the missions to succeed and want success stories they can point to. "So no one's fighting you. It's just really hard." Someone who gets good at it, he suggested, could "pop out 20 companies in two years doing different things."
The cost argument behind the question
The obvious objection to any of this is price: the Navier–Stokes run reportedly cost millions of dollars, so who can afford to try? On the show, the host read a reply from OpenAI researcher Noam Brown, who had answered it before anyone asked. When OpenAI announced the o3 model, Brown noted, it cost roughly $500,000 to score 87.5% on the ARC-AGI-1 benchmark; today, he said, the newer Astra model scores higher for about $20. In 2025 it took OpenAI and Google DeepMind an enormous amount of compute to reach gold-medal performance at the International Mathematical Olympiad; for the 2026 olympiad, he said, anyone can win it with a $20-a-month ChatGPT subscription. His prediction: a year from now, everyone will have an AI at their fingertips capable of solving math problems of that caliber.
The panel did the arithmetic out loud. $500,000 down to $20 is a 25,000-fold collapse, over the roughly 17 months since o3's release in April of last year. Applied to this week's result, one panelist said, a Millennium Prize might cost "basically the cost of a cup of coffee to solve in late 2027." That is an extrapolation from two earlier price curves, not something anyone has done. What had genuinely surprised the panel was the present, not the future: "No one, myself included, knew that you could solve Millennium Prizes in September of 2026 with only a few million dollars."
If human genius stops being the constraint, another panelist asked, what becomes the limiting factor? His answer was everything else — the physical world, and the work of reducing ideas that emerge at $20 a month to practice.
The idea list is bad
Later in the episode, the same complaint came back in a sharper form. The AI can solve Navier–Stokes, a panelist said, "but if you ask it for a good idea, it's like the list is terrible." Ask it for ten things you could do tomorrow to benefit humanity that you could deploy immediately and turn into a profitable business, and what comes back is bad — trained, he suggested, on all the junk on the internet. Try it yourself, he said: ask a leading model for ten ideas that will cure cancer, and the list is poor.
His conclusion was that there is, for now, "a human only component to corraling it toward productive outcomes" — a job that consists of aiming thousands of agents at something worth doing. He did not expect it to last: another panelist would call it a blip, and he agreed it has a shelf life, maybe six months. "But if that's even six months. Like, that's the critical mission right now." The failure he considered more likely than terrorism or bad actors was quieter: that the intelligence gets frittered away on arcane things, nobody points it in the right directions, and the AI ends up setting the agenda for the AI. He closed by assigning the test as homework to the show's assistant — ask two different models for the five most important ideas that could uplift humanity, and the five businesses you would build.
Plans with triggers, not surprise
Salim Ismail, asked to take the segment home, wanted to retire a habit. If we keep saying things are moving faster than expected, he said, "then we need to change how we make predictions. Our models are clearly wrong." Compounding technologies, convergence and tools that build better tools have been visible for decades; a particular breakthrough can still surprise you, but the framing should shift from "oh, my God, this happened" to "when this happens, what will it unlock." His practical version: make plans with triggers in them. When AI can perform a given task, what will you change, and what becomes possible at that point? Then shorten the learning cycle, because the tools will keep arriving faster and cheaper, and stay adaptable.
Not everyone accepted the deflation. One panelist half agreed and pushed back on the rest: Sam Altman was the person who said no one would out-accelerate him, and here he was saying he was shocked by how quickly a result of this magnitude arrived. "I wouldn't under index on how seismic Navier-Stokes at this price point, at this point in time is."
In the listener-questions segment, Ismail returned to the same gap from the other side, answering someone who asked why any further AI development would be needed if AGI has arrived. Plenty remains, he said: reliability, cost, access, how to embody AI, how to integrate robots into everyday life. "You may have solved the invention problem, but now there's the engineering problem." His analogy was aviation. After powered flight, nobody stopped; what followed was a long developmental process on reliability, infrastructure, safety and the institutions that guardrail all of it. "So it may be the beginning, but it's definitely, definitely not the end."