Are We In an A.I. Bubble? I Suspect So.

But if we are, what should we do about it?

Let me preface this post with a disclaimer that I’ve made before: I am a perennial skeptic of pretty much all tech, which is largely a product of my personality. For one thing, I’m good at seeing what might go wrong and less good at seeing incredible upside potential. For another, I’m a consistent late-adopter when I’m not a never-adopter. And once I’ve adopted a particular technology or application thereof, I tend to stick with it long after it has been eclipsed by something clearly superior. There’s a place in finance for someone with my natural inclinations, and that place is the credit markets; I’d make an absolutely terrible venture capitalist. That’s something I’ve demonstrated both by losing money and, more notably, missing great opportunities to make money.

So my attitude toward the LLM generation of artificial intelligence has been unsurprisingly moderate at best. I’ve looked forward personally to self-driving cars—one technology that I was overoptimistic about, at least in terms of timeline—along with automated translation services, document summarization, and other applications that struck me as both plausible and potentially useful to me personally. I’m convinced that artificial intelligence has the potential, once the kinks get worked out, to successfully automate all sorts of routine white-collar work (which it is already starting to do despite the kinks not being worked out yet), and is likely also to speed up incremental improvements in manufacturing as well as a wide variety of kinds of scientific research. I do worry a lot about various downsides—the way A.I. will super-charge fraudsters and other bad actors, the way “good enough but not actually good” A.I.-created art and music will crowd out the real thing, the incredibly destructive effects it is already having on education, and the potential for significant social dislocations. But all of that is par for the course for a “normal” highly disruptive technology.

What I’ve never believed in is super-A.G.I., or even in ordinary A.G.I. I don’t believe the current generation of technology is progressing toward something as flexible and all-purpose as human intelligence, and general super-intelligence is only possible if you have general intelligence in the first place. So the dueling scenarios of utopia and extinction have never struck me as particularly compelling. Even my moderate view, though, is premised on the assumption that A.I. will continue to advance up the steep part of the sigmoid curve for a while before hitting one or another physical constraint that creates a new inflection point and slows advances further. Yet it seems to me that we’re already starting to hit some of those constraints.

The most obvious one is electricity. America is far more invested in the quest for A.G.I. than any other player, and yet our power grid is significantly constrained, and, as the Abundance folks will explain to you ad nauseam, we’re not doing a great job of expanding it’s capacity, and are in some ways starting to go backwards. But there are other less-well-known constraints out there. We’re already running out of “training data” which means trainers need to use earlier generations of A.I. to create additional data on which to train the new generation. It should be obvious why this would be unlikely to lead to advances in capabilities that scale with earlier advances based on real-world data. Then there are the chips themselves. The demand for compute is now growing twice as fast as Moore’s Law, which means that demand for physical chips is doubling on the same time scale that computing power per chip doubles. We used to worry about an A.I. accidentally becoming a paper-clip maximizer, but if we don’t actually reach A.G.I. fairly soon we’ll have to worry about the entire economy turning into a chip maximizer. Meanwhile, we do still need significant advances just to get to my moderate scenario, because actually-existing A.I. is not yet delivering productivity gains, even in the areas, like coding, where it is most optimized to do so.

None of that means that A.I. won’t continue to advance. It just means that we may be in the phase where accelerating investment in A.I. is driven less by accelerating advances that everyone needs to jump on before they get left behind, and more by the fact that those advances are decelerating, so that players who have already invested at an extraordinarily high level have to invest more and more to have a chance of hitting their previously-established goals and timelines. That’s a classic characteristic of a bubble. Indeed, we may even be past that and into the phase where players are playing finance games to artificially inflate a bubble they have come to depend on. See this post from Michael Spencer about how the major players in A.I. have created what looks like a Ponzi scheme, with OpenAI paying Oracle for computing services, Oracle purchasing chips from Nvidia to provide those services, and Nvidia investing in OpenAI, each to the tune of twelve figures. There’s nothing unethical about that kind of daisy chain, but it does goose the numbers for all three companies even though no revenue is actually coming from the end users in the economy for whom all of this capacity is (supposedly) being built. It’s hard to believe that those companies’ CEOs are unaware of that fact, and that’s why suspicion is warranted.

So are we in an A.I. bubble? It sure looks like it to me. That doesn’t mean we won’t get large economic advances (and disruptions) out of A.I. It doesn’t even mean that we won’t ever get A.G.I. or super-intelligence, if it turns out that such things are possible and that the current technology path ultimately leads that way. But I suspect that before we get to wherever we’re going on the technology front, we’re going to hear a big, ugly pop. Which raises the question: if that’s correct, what should we do about it?

My unsatisfying answer is: probably not that much—at least not directly. Alan Greenspan’s wisdom in the face of the dot-com bubble was to say that it’s impossible for policymakers to determine definitively whether we are in a bubble or how big it might be, and that anything the government did to try to burst the bubble would be counterproductive. If there is a bubble, and it eventually bursts, the Fed should respond by cutting rates to keep the economy from tipping into recession, and if it undershoots in its response it should commit to making up for the lost ground swiftly (though implementing the latter without triggering unwanted inflation requires the Fed to corner harder in both directions than it has historically been willing or able to do). But until that happens, the Fed should hold its fire.

I suspect that’s all still good wisdom. I’d also add that any government that set out deliberately to slow growth in the most vibrant sector of the economy would get its head handed to it by the voters—and if the Federal Reserve did it, that would supercharge existing efforts by this administration to threaten its independence. That political analysis applies equally well or better to attempts to regulate A.I. itself on safety or other grounds, which could well have the effect of causing the bubble to burst. Such efforts may or may not be wise, but they are much harder to implement in the middle of a bubble than before the bubble begins to inflate—and if they were to burst the bubble that would be a side effect that would cause them to be discredited which, if you actually believe such regulation is necessary for its stated purpose of safety, is not the outcome you’d want.

If the A.I. bubble were largely equity-financed, I’d leave things there. If it isn’t, though, we have to look at the financial regulatory side of the equation, because the way you get financial crises is when the banking system is threatened systemically by being over-leveraged to a financial asset that crashes. In the dot-com bubble, there was limited leverage deployed and limited banking system exposure. In the sub-prime mortgage bubble, by contrast, there was substantial leverage deployed, frequently in ways that were invisible to regulators. Which is the A.I. bubble more like?

I would have guessed that it was more like the dot-com bubble, but this Noah Smith post convinced me that there might be invisible leverage to worry about. In a nutshell (you really should read the post if you are interested in the details), most of the financing of the high-flying tech companies is indeed in the form of equity, but a lot of the financing of data centers is done through asset-backed loans which are securitized and then insured. I don’t know anything about the models used for those securitizations, but the key component that can go wrong is the expected correlation between different assets in a pool. When an asset bubble bursts, that correlation can quickly get very high, and if that possibility isn’t properly built into the modeling (as it wasn’t with respect to sub-prime mortgages), then entities like banks and insurers who think they are holding negligible tail risk can suddenly face losses that are many multiples of what they could actually survive. And then you have a financial crisis.

Could regulators do anything about this? Probably. The thing to focus on would be whether banks and insurance companies are adequately capitalized against the risks they are actually taking. If they aren’t, based on the regulator’s own assessment of the risk in the positions, you make them hold more capital. That’ll reduce the profitability of those positions, which will encourage the banks and insurers to unwind a portion, or at least not add more, which in turn would force data center builders to find other sources of financing. That, potentially, could burst the bubble, but it might not—A.I. companies might be able to keep going with more exotic but less systemically-threatening forms of equity financing—and if it did burst the bubble the regulatory move is sufficiently distant from the outcome that you might have a better chance politically of surviving than you would if you tried to burst the bubble directly by hiking rates or if you tried to do it indirectly by imposing A.I. safety regulations. Moreover, I suspect that if there’s a real problem growing in their portfolios, banks and insurers might themselves be quiet allies of a sensible regulatory effort. This is precisely the kind of situation where an industry will be unable to self-regulate even if it sees the potential for catastrophe, because competitive pressures will force all players to take advantage of lax regulation lest they lose market share, but the amount of profit at stake for the banks and insurers themselves can’t be anywhere near where it was in the subprime mortgage days, so we’re probably not at the point yet where that business is too big to be constrained. I’m too far away from finance these days to have anything like an informed opinion on how worried anybody should be, but if we should be worried specifically about the financial sector, there are likely things regulators could do to get ahead of the problem.

But I feel confident in saying that if there is indeed something to worry about, this administration is neither going to know nor take any action to address it. That fact should probably also figure into everyone’s assessment of how likely we are to be headed for trouble.