Economic Lessons from the Last Few Singularities

One of John von Neumann's absurdly varied contributions to human knowledge was the idea of a technological singularity, i.e. a point at which advances in technology and economic growth happen at such a fast pace that prediction is impossible. This doesn't mean it's the end of history, and in fact means that history happens faster than before. It is, however, the end of the kind of history that anyone in the past could extrapolate from. We've had them before: agriculture, writing, animal domestication, and early forms of government were one such qualitative change. It's tricky to put exact dates on when the rise of manufacturing led to this kind of change, but certainly the Second Industrial Revolution counts—reducing the cost of transportation and communication, mass-producing goods, and finding scalable sources of fertilizer raised the maximum sustainable human population by a bit and their collective energy output by a lot more. The mid-twentieth century, representing both the peak deployment phase for earlier industrial technologies and a thus-far-unmatched era of sustained productivity growth also counts. And within the realm of information specifically, arguably the combination of Moore's Law and the growth of telecommunications created something similar—even if someone half a century ago could imagine something like watching a movie on-demand on a plane, what you probably wouldn't have been able to imagine is just how many products would be available with same-day shipping, or how trivial it would be to track down obscure information.

Now's an interesting time to think about this, because we probably aren't in such a singularity just yet, but are at the point where there's suddenly more scope to keep investing in AI, so the possibility is there. A few months ago, there were concerns about hitting a data wall, and GPT-5 is behind schedule and not performing as well as expected, at least according to the WSJ ($). But o3 was really impressive, both for the capabilities that it demonstrated and for how it got to them. The short summary on both:

A lightweight version of the model scored 75.7% on the ARC evaluation, and the full model spent quite a bit more to score 87.5%. Two years ago, François Chollet speculated that a model would clear the 70% threshold in 2026-2028.

Its approach involved scaling test-time compute, i.e. generating a bunch of plans, iterating on them, and eventually converging on the right approach and the right answer. Specifically, the smaller model generated 33m tokens, and the larger one 5.7bn. (For context, the full transcript of the lightweight model was roughly 44x the length of War and Peace.) Scaling data and compute to improve models has worked for a long time, but scaling compute along to get better performance. This thread is illuminating, in that the author, a math professor, says that he knew how to solve the problems close to his domain and would expect an expert in other domains to be similarly capable of figuring out an approach—so, not superhuman, but good enough that we're grading it based on comparisons to people who've been selected for generally relevant skills and who specialize withing their category.

(It probably goes without saying, but adding a new dimension to scaling—run the model longer to eke out IQ points—means there's a broader set of economically valuable existing activities that these models can address, and also a broader set of new services that will come into existence once expertise-as-a-service has been available for a while.)

It's a meaningful and hard-to-predict change if most of the thinking that happens on behalf of people is happening in silicon rather than in wetware, just as it was a big transition when the marginal unit of energy used to grow food switched from humans to draft animals, and the later shift where the marginal unit of energy switched from currently-living things to hydrocarbons.[1]

Technological singularities usually entail more positive liberty and less negative liberty. That's always been the case. Hunter-gatherers don't need much of a concept of ownership beyond their personal possessions. An independent craftsman can set his own working hours, albeit within the confines of the work that's required to stay alive; a factory worker needs to clock in and out. But an economy with agriculture can support a larger population, and generally needs to support more complex institutions.[2] An AI-heavier world is one where it makes economic sense to collect a lot more data, and to incorporate it into models or automatically detect anomalies (if your smart glasses are connected to a social network that also controls a lot of purchase data, maybe your Christmas shopping in 2025 will consist of wearing them while wandering around different stores while the model reminds you of which friends and family members would like—but haven't yet purchased—some product that happens to be within your field of vision). That also means it's a world where you're being watched more closely, albeit by algorithms rather than conscious observers, and where deviations from expected norms are easier to detect and punish.

Singularities change which questions it makes sense to ask; they mostly eliminate some and they make others more salient. To a hunter-gatherer, a question like "Where will your ancestors live" is incoherent; if they exist, they'll be wherever the hunting is good and the berries are abundant, and when that changes they'll pick up and move somewhere else. In an agrarian system, the question is often not worth answering, because the answer is "right here, where the farm is, because otherwise they'd starve." Now it's a statistical question—probabilistic thinking being the sort of thing that makes a lot of sense in what an agrarian would consider our post-scarcity society, which isn't dominated by binary issues of life and death but is full of uncertain tradeoffs.

And there are some things they don't change. The ranking of biggest cities fluctuates, but their advantages compound over time; ports, administrative centers on rivers, cities centrally connected to road networks, etc. tend to endure—in fact, the centrality of these networks is reinforced when transportation throughput goes up, because they have the relevant infrastructure already and its rising opportunity cost prices out the least efficient uses.

Some institutions seem to last through this kind of transition just fine, typically when they combine some sort of natural long-term risk-aversion with an understanding of the difference between the service they provide and the exact process they go through. A well-born young person today can go to the same university, for the same reason, that someone similarly situated did 900 years ago at Oxford, 700 years ago at the Sorbonne, or 300 years ago at Harvard. There are centuries-old banks and insurance companies, albeit often under newer corporate umbrellas; a financial business can change many of the details of its day-to-day function while retaining the same fundamental form. Hotels, too, can last a surprisingly long time; they're the kind of central institution around which other businesses accrete over time, and they end up capturing some of that upside for themselves. What these businesses have in common is that their outcomes are more driven by downside than upside; a successful bank or insurance company is one that didn't make whatever big mistake its peers did.

We are headed for a weirder world in some respects, and there's a paradox to the bundles of technological and institutional changes that lead to this kind of shift. They tend to start with repeated productivity gains in a narrow sector of the economy, but these gains are so sustained that the relevant part of the economy end up intersecting with the rest of it; there aren't really "Internet companies" today any more than there are "electricity companies" or "sectors of the economy where literacy is beneficial." But that's also one of the constraints on deployment; none of these technologies were economic pixie dust that could automatically improve whatever they were applied to, and the biggest winners from them tended to think about applications rather than underlying products—Henry Ford made his fortune from the internal combustion engine, but he was fundamentally obsessed with cars more so than particular propulsion techniques, and explored electric vehicles early on.

Calling it "the singularity" was always a misnomer, because they've happened before and don't look that way in retrospect. Those previous examples did lead to radical change that contemporary observers couldn't possibly have comprehended, but enough stayed the same that even though the future wasn't comprehensible to them in the past, that past is reasonably understandable to us. Perhaps they don't count as singularities because they eventually leveled off—you can't surpass 100% literacy or urbanization, and per-capita energy use in the developed world mostly leveled off in the 90s and has declined since then. But even if that happens to AI—even if we build a model slightly smarter than the smartest person and it uses its intelligence to definitively prove that nothing can be smarter than it, we're still in a weird and unpredictable situation, and after a while we'll probably start poking around for the next singularity.

If you want to situate o3 in this model, taking the optimistic view that AI happens as its advocates say it will, I'd compare it to something like oil and natural gas in the context of a coal-based energy system—i.e. an improvement that changes the context in which a broad concept can be employed, but less of a step-function than an acceleration of an existing trend. ↩︎

I'm in the middle of an enjoyable history of the Dutch Republic right now, and one thing the book points out is that Holland needed more complex and more local institutions earlier than other parts of Europe, because they had to reclaim land and manage the dike system. It's hard to do that from an imperial court hundreds of miles away, so they had to figure it out on their own. Incidentally, medieval Holland was also a very early example of the export-driven economic development model of accumulating capital equipment that can process raw materials into high value-added exports. In their case the input was Baltic grain, the capital was cattle, and the exports were cheese. ↩︎