There’s this old saying about project management: You can have a project done (a) fast, (b) well, or (c) cheap. Select two of these three options.
(fast+well, but $$$), or (well+cheap, but slow), or (fast+cheap, but shoddy).
This sprung to mind today while reading Ethan Mollick’s latest essay, “The Dot and the Swarm.” Mollick is convinced that AI’s “Bitter Lesson” has once again blown through all expectations. (roughly: the “bitter lesson” states that there is no AI problem that cannot be solved by throwing an order of magnitude more compute at the problem. Scaling rules all.) AI agents have solved intractable management problems. OpenAIs new Dot and Meta’s Muse are the ultimate personal assistants. Swarms of AI agents are now solving Millennium math problems. Just keep adding compute, and every barrier falls away.
I don’t want to dismiss this out of hand. The latest wave of agentic AI can, indeed, do a bunch of stuff that the previous wave couldn’t do. I think it’s important for skeptics like me to take that seriously. I haven’t tried Muse or Dots. I’m unlikely to try either one, for the same reason that Ryan Broderick articulated in his newsletter last month. There isn’t actually anything in my day-to-day life that I want a robo-assistant to tackle for me. Let’s grant the possibility that Muse/Dots will be as potent a tool for budding entrepreneurs as Claude Code has been for coders.
But even so, I find it striking that the way OpenAI pulls off these Mollick-astonishing feats is by spending truly immense sums. Mollick writes “[Navier-Stokes] is among the most famous open problems in mathematics, with a $1 million prize, but OpenAI apparently solved it using AI alone in 88 hours.” It is, of course, much more complicated than that. But what he particularly fails to mention is that OpenAI devoted 88 hours and 130 billion tokens to the task. All told, the company spent somewhere between $10 million and $40 million to solve Navier-Stokes in 88 hours.
This seems pretty relevant. It leads me to the following four observations:
(1) It is indeed impressive that frontier LLMs, given infinite budgets and access to all of the latest mathematical research, can generate this proof. OpenAI’s previous model could not accomplish this benchmark feat. It ain’t nothing.
(2) $10-40 million is not a lot of money for OpenAI right now.
(3) $10-40 million is a lot of money, relative to the amount being spent on pure mathematical research right now.
(4) OpenAI isn’t in this for the pure math achievements. It also isn’t in it for the prize money. (I’m no economist or mathematician, but I think I can spot the problem with spending $10 million to claim a $1 million prize.) OpenAI is in it for the publicity. Solving Navier-Stokes helps OpenAI maintain the air of futurity. Silicon Valley runs on futurity. Spending limitless money to solve a famous math problem leads to fawning media coverage which in turn leads to the next round of limitless investor cash. It’s a cycle that cannot last forever, but it sure seems otherwise when you’re in the middle of it.
We are not heading for a future where LLMs solve all the math problems and mathematics as a professional field ceases to exist. We are likely heading for a future where LLMs are incorporated into mathematical research. And the boundaries of that future will be dictated by scarce resources, and by the choices made by the people who control those resources.
The same is true about Dots and Muse. OpenAI and Meta are operating these products at a loss, throwing tremendous money at the project. When you have infinite money to throw at a problem, you can move fast and produce something that works quite well. (Though the security issues are, uh, ongoing. Maybe don’t sign up to be an unwitting beta tester for these things?) This is deceptively impressive. It doesn’t give us much insight into how the product will work at scale, what sort of resources it requires in order to maintain quality outputs, or what the profit-generating market for it will be.
This is one of the themes in Vaporware, Inc (which, btw, you could absolutely preorder today.): of all the potential, imaginable futures of an emerging technology, the future bends towards money. The optimistic tech futurists always call attention to how a technology might be used, while ignoring market incentives that are easily predictable in advance. Pay attention to the money, and you’ll have a much better grasp of where the future is headed.
One of the defining qualities of the current AI race is that the frontier labs are all operating as though they have infinite money to spend. This, I think, is what’s driving some of their most reckless behavior. News broke a couple weeks ago that Anthropic has set up its own wet lab. (This briefly drove me insane. Dario Amodei both thinks we should collectively slow down the pace of AI development to prevent global disaster and ALSO thinks his company should give Claudebot its own supervirus-production facilities as a treat.)
The simplest explanation for why is Anthropic rushing to give its AI model the capacity to produce novel viruses is “because Anthropic has infinite money, and needs to maintain a steady cadence of futuristic production.” Maybe the AI model speeds up drug discovery, cures some rare diseases, and disrupts all of pharma. Given time and infinite resources, who knows what Claude might do?
Giving infinite money to a handful of companies and encouraging them to outdo each other in an all-against-all race is, well, bad. That is a very bad way to build the future, and we should not be doing so.
The answer is to rein in their resources. Let them build agentic AI while facing typical resource constraints. The result will be companies that build slower and make fewer catastrophic errors. The result will also be companies that don’t build their own wet labs on a lark because those companies lack the expertise and qualifications to get into biological research, and they don’t have infinite money to throw at futurish ideas that sound cool/potentially catastrophic.
In the meantime, I’m going to assume a position of guarded skepticism whenever the AI industry announces a new benchmark or breakthrough — not because the benchmark achievements are necessarily fake, but because companies chasing futurity while operating with infinite money don’t give us much insight into what the real-world trajectory of their products is likely to be.