Artificial intelligence and Engels’ Pause

{"text":[[{"start":6.15,"text":"Ryan Avent is a journalist, author of In Good Faith, and writes a newsletter called The Bellows. "}],[{"start":12.3,"text":"How exactly is AI reshaping our economic prospects? Are we in “the foothills of the singularity”? Are we facing an imminent jobs apocalypse? Will the bursting of the market bubble send the economy into an economic tailspin? "}],[{"start":26.450000000000003,"text":"The cone of outcomes certainly appears to be very wide."}],[{"start":null,"text":"

"}],[{"start":31.050000000000004,"text":"In recent months, as businesses have abandoned “tokenmaxxing” in the face of higher user costs, opinion seems to have shifted towards the view that AI might be less impressive than expected, with the implication that the capex boom may be running out of steam. Paul Krugman, for instance, reckons that:"}],[{"start":48.2,"text":"[T]he surging demand for “compute,” which has been used to justify huge spending on datacenters, is clearly slowing and perhaps even cresting . . . "}],[{"start":58.050000000000004,"text":"[W]hile AI may yet prove to be a strongly capital-biased, wage-depressing technology, the evidence so far is consistent with a more modest, less apocalyptic scenario — more consistent with such a modest scenario than it seemed to be even a few months ago."}],[{"start":74.45,"text":"Less apocalyptic is good. But is cresting demand for compute the right reason to expect a more economically modest impact from AI?"}],[{"start":83.3,"text":"Back in April companies like OpenAI and Anthropic raised user costs, forcing the businesses that were using frontier models to write out-of-office messages to assess what value they were getting for their money. The answer, in many cases, seemed to be: not enough. The story thus looks like one in which AI companies are jacking user fees to cover otherworldly levels of capital investment, even as businesses judge that AI is no great shakes. That would be . . . bad for the boom."}],[{"start":110.05,"text":"Yet it could be that something else is happening. We know that frontier models are really good. Claude Opus and GPT-5 wowed users, triggering the rush to tokenmaxx, and Claude Mythos is too scary good to be released to the public without major restrictions. "}],[{"start":127.75,"text":"Tokenmaxxing isn’t the only thing driving an explosion in token consumption this year; businesses really are incorporating highly capable models into their work processes. We also know that the AI labs have been woefully unable to meet user demand. Over the first few months of this year, users complained that Anthropic dumbed down Claude during peak periods, and the number and length of Claude outages increased. In addition, it is widely assumed that Anthropic limited initial access to Mythos because the computing resources to meet expected demand simply weren’t there."}],[{"start":null,"text":""}],[{"start":162.4,"text":"Rationing demand by raising prices is a natural response to this computing crunch. Higher prices crowd out low-value AI activity, and there is plenty of that around. But the users getting good value out of AI should be willing to pay the higher prices, boosting AI companies’ margins and helping to fund ongoing capex. "}],[{"start":181.5,"text":"This flywheel could spin indefinitely, both because more compute will mean more capable models, and because any easing of AI costs associated with increased capacity should quickly unlock newly economical use cases."}],[{"start":197.15,"text":"We’ve seen something like this before. The new technologies of the Industrial Revolution created fabulous new production possibilities. Unlocking them, though, meant engaging in a lot of complementary capital investment: in mechanical looms and mines and railways and such. "}],[{"start":213.70000000000002,"text":"The economic historian Robert Allen argues that soaring demand for scarce capital drove up profit rates for capital owners, and these profits were ploughed back into new investment. For the better part of a century, capital’s share of British national income marched upward, from about 20 per cent in 1770 to nearly 50 per cent in 1870."}],[{"start":null,"text":""}],[{"start":234.00000000000003,"text":"Unfortunately, this was the dynamic behind what Allen termed “Engels’ Pause” — the fact that over most of the first century of industrialisation, output per worker rose steadily even as real living standards stagnated. "}],[{"start":247.85000000000002,"text":"As Friedrich Engels observed in The Condition of the Working Class in England, first published in 1845 a few years before co-authoring the Communist Manifesto with Karl Marx:"}],[{"start":261.1,"text":"Since the Reform Act of 1832 the most important social issue in England has been the condition of the working classes, who form the vast majority of the English people . . . What is to become of these propertyless millions who own nothing and consume today what they earned yesterday? . . . The English middle classes prefer to ignore the distress of the workers and this is particularly true of the industrialists, who grow rich on the misery of the mass of wage earners."}],[{"start":287.75,"text":"Workers lost out, on the one hand, because scarce capital captured an outsized share of the gains from growth, and on the other, because high financing costs meant that investment in things like housing and public works had to be done on the cheap. Only in the last few decades of the 19th century did the returns to growth become more balanced between capital and labour, such that wages began rising alongside productivity."}],[{"start":310.5,"text":"The parallel with AI isn’t a perfect one (after all, a lot has changed over the past quarter millennium). The deployment of ever more powerful AI everywhere isn’t a sure thing. The pace of model improvement could flag, or governments could decide that the best models are too dangerous to be used in the wild. Achieving the right balance between supply and demand is also tricky, and price hikes could be constrained if AI companies are reluctant to cede too many users to rivals. "}],[{"start":337.25,"text":"But if AI does transform the economy, the levels of investment required will be unlike anything anyone living has ever seen. If that is the case, then what will the next decade or two look like, economically speaking? "}],[{"start":351.15,"text":"Fortunately or unfortunately, the singularity probably isn’t near, because AI can’t do every little thing until a truly epic amount of expensive and time-consuming investment has taken place: in data centres and power plants and chip fabs and robot factories and apocalypse bunkers. "}],[{"start":369.84999999999997,"text":"More encouragingly, it also suggests that a jobs crisis is further away than we think, because rationing of scarce compute by price will keep workers cost-competitive for years to come. The rub is that labour will lose out in other ways — from downward pressure on wages and inadequate investment in non-AI things — even as the rich get much, much richer."}],[{"start":391.7,"text":"If that feels like the makings of a popular revolution, well, it isn’t called Engels’ Pause for nothing. In the end, the biggest threats to the AI boom are likely to be social and political rather than economic."}],[{"start":410.4,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1782726421_8361.mp3"}