Nine Hundred Occupations, Twenty Years, and Still No AI Revolution

Labenz invites Hanson to sketch the big picture that runs through his work: modern humans live in a strange dreamtime, unlike the world of those who came before us and unlike the world of those who will come after. Hanson starts with growth. We are in a period of very rapid change that simply cannot continue for long on a cosmological clock. Ten thousand years, he says, would already be far longer than this pace could manage. When growth slows, population can expand relative to the economy, and the world moves back toward a Malthusian condition of the sort almost all of our ancestors lived under until a few hundred years ago.

The other anomaly is integration. Distant ancestors were culturally fragmented; each person talked to a small group nearby. Distant descendants will be fragmented across the universe, unable to speak instantaneously over light-speed distances. Only now can an entire civilization talk with a communication delay that is tiny compared with the doubling time of a fast-growing economy. We are rich, globally wired, and under little pressure to do what it takes to survive. That, Hanson says, is why we can indulge delusions, emotions, and dramatic visions. Our minds were trained in a harsher world and have not been re-disciplined for this one.

Labenz tries a punctuated-equilibrium frame: a Cambrian-style interval in which constraints briefly relax before natural selection reasserts itself. Hanson rejects the analogy. After multicellularity, evolution had a stable new opportunity to adapt to, and then it settled. Humans today keep getting new shocks. There was not one change three hundred years ago that we are slowly fitting ourselves to. We will not see slow adaptation to a new thing, he says, until we actually get a stable new thing — and we have not.

He is just as unwilling to lead with values. Most grand talk, he argues, is a hunt for value buddies: people skip the armaments and the budgets and jump to which side you are on. He is literally afraid that if he states his values, the conversation will go there and never leave. The one preference he will own is that humanity should keep growing and take a big chunk of the universe. A million years ago, people who had been able to freeze their familiar world would have blocked the path to us. If we want scale, he says, we will have to give a lot on how similar the future is to our own.

Hanson’s 2016 book *The Age of Em* is built on one technical assumption. Scan a particular human brain in enough spatial and chemical detail to map cells, synapses, and connections. Fill in a computer model for each cell. If the map and the cell models are good enough, the input-output behavior of the simulation should match the original brain. Hook it to artificial eyes, ears, hands, and a mouth, and you have a substitute that can work throughout the economy. The point of the book, he says, was to show that a specific assumption can be worked into a large set of consequences — the kind of detail most futurism never attempts.

Anything stored as a computer file can be copied cheaply, paused, archived, run faster or slower, and moved at the speed of light. Hanson is not assuming that ems are magically transparent. Human brains are a mess, and so are large legacy software systems: you can inspect a piece, but you mostly have to take the whole thing as a unit and make only modest changes. That is close to what he assumes for ems. Labenz notes that a few hundred base ems might cover the mental styles of roughly 900 human occupations, with cheaper variants filling in the rest — a structure that already looks a bit like GPT-4, Claude, Gemini, and a swarm of fine-tunes.

A large language model is not yet a full human substitute. Hanson’s first fork is therefore brutally practical: are we forecasting something close to today’s models, or a descendant that can stand in for people across a wide range of jobs? Stick with current systems, he says, and the exercise is closer to forecasting a microwave oven — where it helps, what it displaces, what complements it. The Age of Em was a grand exercise because ems change everything. Most technologies do not.

Labenz wants a fourth category on the table: Mamba-style state-space models, which keep a fixed-size internal state that evolves through time and might support something closer to employee-style onboarding. Hanson has a problem with that fourth box. Across the history of computers, and of automation before them, a wide range of paradigms has been tried. Each time a new structure appears, it comes with a demo that prior systems could not do. People get shocked, then they ask whether we are almost done. They said that centuries ago and half a century ago. The prior that the next exciting paradigm is the last one we will need is, he says, pretty low.

Hanson was a full-time AI researcher from 1984 to 1993 at NASA and then Lockheed. He will talk about what language models can do in the near term. He will also talk about the later world in which machines can do pretty much all human tasks, whoever gets us there. What he will not grant is that we are close to that point, or that most jobs will be replaced in a decade or two on the current approach. When forecasters are asked when most jobs go, he notes, they usually put the date far out — because they do not actually think we are close.

The economic test he wants is simple. Treat the model as a new employee: watch how I do this, now you do it. That, he says, is the fundamental task in the economy. Today’s language models are not up to it across most jobs. Labenz cites the GPT-3 paper’s finding that large language models are few-shot learners, and the medical proverb of watch one, do one, teach one. Hanson does not deny that you can say “here are some examples, give me another.” He denies that the *range* is wide. Most people, on most of their main tasks, could not swap a language model in that way.

Labenz keeps returning to a qualitative jump: these systems do something earlier AI could not. Hanson’s reply is that every previous generation had that same conversation. There are apparently a lot of things near the core of what it is to think. Ilya Sutskever’s line that he can speak to a model and feel understood is, Hanson grants, a real shock. What he denies is the leap from shock to “we must be near the end of the path.”

Labenz offers a threshold he thinks has just been crossed. A Google DeepMind paper, on cases lined up as if you were chatting with a doctor, showed a language model at about 60 percent diagnostic accuracy against 30 percent for human doctors, with human-plus-AI in the middle. Hanson is unmoved. One reason he entered AI in the early 1980s was a wave of journal articles and national media coverage claiming that the best systems of the day — expert systems — already did human-level medical diagnosis. Either those studies were cooked, he says, or the ability was real and medical licensing forbade applying it. The legal barrier is still there.

Labenz argues that today’s chat interface is qualitatively different: a patient can talk in ordinary language instead of walking a wizard and punching numeric scores. Hanson says that was not the limiting factor. You could have made 1980s expert systems usable by ordinary people with modest effort. Only doctors are allowed to give medical diagnoses, so only doctors are allowed to use the tools with patients. Fancy graphics were missing then too. The constraint that matters is still the law, and it does not look about to vanish.

He reaches for nuclear power. The world has long believed nuclear energy is powerful; that belief is exactly why it was regulated into a corner, and cheap energy never arrived. There is a risk, he says, that we do the same with AI: convince people it is so powerful it must be locked down. Labenz has been encouraged by the medical establishment, including Harvard Medical School professor Zak Kohane, who had early GPT-4 access and wrote a book timed to the launch. Hanson splits the cases. Internal use by doctors has always been allowed. The thing that never scaled is that doctors did not want to type more into the computer during short visits. Computer diagnostic aids that would have improved accuracy at the cost of more data entry have been available for decades, and doctors have declined to spend the time.

The same licensing logic, he adds, already wastes a cheaper human substitute. Randomized experiments decades ago found that nurse practitioners and paramedics did just as well as physicians at first-level general practice, at perhaps a factor of two or three lower cost. They have not been allowed to take that role. If ems are invented, Hanson does not assume eager adoption. Resistance is more likely, then a dam breaking: an ambitious country — he mentions North Korea — going whole hog once the technology is cheap enough. The world might coordinate to hold it back for a long time, but not a thousand years. A sudden flood, he says, could mean a lot of unhappy people, and maybe a lot of dead people.

In the months before the taping, Hanson says he dramatically changed his picture of the future. Rising population will stop and then fall with fertility. Innovation will grind to a halt. World population will keep falling until insular, high-fertility subcultures such as the Amish grow from tiny bases into the dominant population. When that civilization is large enough, innovation turns on again, and then AI and ems restart. The world that eventually produces them would be very different from ours, and hard to think about. This is not what he wishes. He would rather see the growth rate of the past century continue for a few more.

He has tried to put a number on the lost progress. If we had stayed on a steady growth path, the total innovation over that multi-century pause is, in his estimate, about 60 to 90 years of ordinary progress. If we can get full human-level AI in that window, population decline matters less: AIs take over most jobs and the world economy can keep growing. He thinks that is iffy. Many people are sure it happens in the next ten years, and therefore sure it happens in 60 to 90. Hanson looks at the history of AI excitement and hears a familiar song.

Even a large increase in automation might not be enough. If the next three 20-year periods each move automation by another third of a standard deviation, 60 years would shift the distribution by a full standard deviation. That would mean a more productive economy in which humans still collect most income. It would not, on his account, be enough to stop the economy from shrinking as population shrinks. The well of further automation would then dry up, because innovation itself would stop. AI is especially at risk in a shrinking world: it is a large-scale, concentrated technology run by very large enterprises, not by mom-and-pops, and scale economies are exactly what a smaller economy loses.

The period from 1999 to 2019 was full of claims that AI was about to revolutionize society. Hanson and coauthor Keller Scholl looked at all jobs in the United States — roughly 900 kinds — with a yearly measure of how automated each job was. They asked whether more automation moved wages or headcount up or down, and whether the traits that predict automation changed, as they would have if the nature of automation had been revolutionized. On average, when jobs got more or less automated, wages and the number of workers did not move. The predictors did not change at all.

Those predictors are old-fashioned. Jobs automate more when they have clear measures of how well you did, when they sit in a clean environment with fewer disruptions, and when nearby tasks have already been automated, because automation spreads through a network of adjacent work. The distribution of how automated jobs are is roughly Gaussian. Over those 20 years the median moved about a third of a standard deviation. Jobs got more automated, but most still are not highly automated. Near the top of the list: airline pilots, who almost always have automation helping. Near the bottom: carpet installers, who staple in carpets with essentially none.

Labenz puts up MMLU, the exam-style benchmark on which GPT-4 sits in the high 80s to 90 percent across fields. Hanson says the benchmark is just wrong as a proxy for the economy. In the 1960s, AI researchers treated chess as a paradigm of thinking: if a machine could play at a high level, it would have to have solved most of the major problems of intelligence. They even forecast, from the improvement rate, roughly when human-level chess would arrive — and it did, in the 1990s, about 25 years before this conversation. Then general intelligence did not follow. There turned out to be a way to be good at chess, and at exam answering, without the full range of abilities the economy needs.

People often say that because compute grows exponentially, economic impact should explode after a quiet start. That is not what the record shows. Hanson’s model is that the computing power required to automate tasks is log-normal with a very large variance. Exponential hardware growth then walks through that distribution in a roughly linear way. Labenz restates it as a low-hanging-fruit picture: the ladder is growing exponentially into the tree, and each extra rung lets you pick more fruit, but it is a really tall tree. Hanson’s forecast for the next 20 years is the last 20 years again: slow gains, mostly basic automation, and little dent from advanced AI in the larger economy.

Labenz is not persuaded. Two years earlier, AI art on Twitter was mostly a curiosity; his teammates at Waymark then made a short film with a half-dozen people and DALL-E 2 at a production quality that, he says, would once have taken a crew in Antarctica. Midjourney looks photorealistic. Deepfakes and cloned voices are close enough that 2024 will be full of scam stories. His own programming productivity with GPT-4 is up several-fold, not incrementally. Protein design and medical diagnosis are moving too. The same relatively simple technology is hitting many industries at once.

Hanson’s reply is a selection effect. In the 1960s a U.S. presidential commission studied whether most jobs were about to be automated. Ever since, periodic articles have picked not random tasks but the places where automation has made the most difference. Image generation is this year’s vivid example, and he is happy to admit it is real. A 20-year-old should go where the excitement is — “move west, young man” — but still ask what fraction of the country the West is. Graphic artists may be hurting. That does not tell you how far the rest of the task list still has to go.

Programming is the biggest profession in Labenz’s list, and here Hanson has a concrete counterexample. His younger son is a professional programmer doing routine application development, not frontier work. The firm looked at large language models to help write software; the evaluation was that they are not useful now, and they will wait six months and look again. Hanson thinks that matches what most working programmers are actually doing. If some people are getting a factor-of-two productivity increase, we should eventually see it in wages. Demand for programming is highly elastic: the world is short of decent programmers, and a lot of money would hire more of them if they were effectively cheaper.

He contrasts that with piano tuners. Tuning is a small part of the cost of having a piano, so demand is inelastic: double a tuner’s speed and you need about half as many tuners. Programming is the opposite. Expand the supply and the price should not fall much, which means you should see a much larger quantity of software *and* higher pay per more-productive programmer. Wages for programmers have already risen a lot over the past decade without a matching flood of new people, because the marginal programmer is hard to produce. If language models were really multiplying productivity, Hanson expects a large further wage increase. He even offers it as a bet over five or ten years. Labenz agrees the quantity of programmers will grow; he is less sure wages will.

On the em analogy, Labenz notes that a handful of base models plus cheap fine-tunes already dominate, much as Hanson imagined a few hundred canonical em personalities covering most jobs. Hanson agrees the structure rhymes. There are about 900 named occupations, but many share mental styles; a few hundred humans might cover them, and the number of base AIs could be four, 40, 400, or even one or two. The harder contrast is merging. To make GPT-4, labs did not continue training GPT-3; they started from a blank network. Nobody, he says, knows how to merge Claude and GPT-4 in a sensible way. Starting from Claude and then doing GPT-4’s training would likely be worse than starting blank. That is a substantial degree of not being able to merge, and it is unlike humans, who reuse old skills to learn new ones.

The deeper issue is rot. Large legacy software rots; human brains rot; old companies rot and get replaced. Hanson expects the same for language models and for whatever AIs we build. It is a reason to doubt the image of one super-AI that lasts forever: that system would rot and would have to replace itself with fresh versions, which could be substantially different. Biology could have made immortal organisms and did not. In a few years, he predicts, a company will have a customer-service model it has been building up, the model will be rotting, and the firm will not know how to make a new one that inherits what the old one was taught without inheriting the rot. That design problem is not a side issue. It is the long-run problem of complicated systems.

Moral status follows from bargaining, not from metaphysics. Ems running a society would insist on weight, the way you insist that people dealing with you give you some. Humans who want to work with them would have to show modest respect, at least in public. The economics of slavery is the guide. Slavery paid when labor was scarce and wages were high; it paid for simple pain-through tasks such as picking cotton, not for house slaves and city shop slaves who had too many ways to sabotage you if they were angry. Work has only gotten more complicated. Language models trained on human conversation have absorbed the human pattern: people resent being treated as slaves, need to feel respected to be motivated, and sabotage less when they have some freedom. In the vast space of possible AIs there could be minds that do not mind enslavement. Language models, Hanson says, are not going to be those.

What would change his forecast? Not demos, and not the high-tech firms that get a reputation for doing AI. He would watch which tasks actually get automated, how much is paid for those substitutions, and whether worker numbers, wages, and the set of supplying firms move. Tractors replacing farmers and trucks replacing horses were large, visible, and obvious. You did not have to scrape for subtleties. Even those substitutions did not mean AI was about to take over everything. But they would tell him what he was looking at. Labenz, introducing the episode, had already conceded part of the gap: an AI educator told him that failure to form new habits is the most common reason AI projects fail; a Microsoft GitHub salesperson once knew less about Copilot than someone on the customer’s own team; and most people were never trained to break their work into the discrete tasks models can handle. Feasibility and adoption are different problems. The disagreement that remains is empirical: whether the efficiency gains ever write themselves into the wage bill.