A $100 Doctor Visit and a $1 AI: GPT-4 Was Already Unbundling Work

On *The Cognitive Revolution*, Nathan Labenz returns with co-host Erik Torenberg to unpack a claim he made to OpenAI while red-teaming GPT-4: there will be economic transformation. The details, he says, are still being worked out. The method he used in those two months of early access was deliberately simple. He asked the model to play the professional advisors people already pay real money for, and he stayed in the conversation long enough to see whether it had substance.

The first role was a doctor. In the United States, a medical concern still means booking an appointment, driving in, and often losing half a day in a waiting room, after which many patients feel they were not really listened to. Setting up an AI doctor, by contrast, was "trivially easy": a simple role-cast prompt, then a back-and-forth that Labenz found extremely easy to replicate compared with visits he had actually had. He is not a physician, and he is careful about evaluation, so he kept the tests few and personal.

The cases were minor issues of his own and worries about his three-year-old that might otherwise have meant a call to the pediatrician. Within those two months the family reached a clear substitution point. One incident produced advice they judged reliable enough to skip the call entirely. It turned out not to be a real problem, which makes it an easy case, but it was still a call that would have been made and was not, because they had GPT-4.

The dentist test was more dramatic. Labenz described a weird spot on a tooth and the model told him the dentist's treatment sounded non-standard, out of line with best practices. That turned out to be true. He also tried immigration law, tech support for his grandmother, and even play with a three-year-old. For a fictional fence dispute between neighbors he asked the model to mediate while he played both sides, then sent the transcript to a lawyer friend. The lawyer said the legal answer was open and shut: the fence was on the other person's land, so that person decides. Labenz found the contrast revealing. In a situation where the neighbors still have to live next to each other, what people actually needed was a mediator, not a ruling, and the role you assign the model changes how useful it is.

A virtual trainer in a group chat handed out specific advice to Labenz (in okay shape), his pregnant wife, and his frequently injured brother, in a tone Labenz said would not survive a double-blind Turing test against a human text coach. A virtual garage dispatcher was more patient than a real mechanic with grease on his hands, because the model does not mind sitting on the phone until the caller is satisfied. A hardware-store associate recommended a specific bulb for 1980s halogen fixtures in Labenz's basement; he searched, ordered, and it fit. A solar consult walked him through charging phones versus running a fridge or an air conditioner, and talked him out of the last of those on a retail panel. His wife, who organizes thousand-person all-vegan events with a rotating cast of caterers, looked at a GPT-planned menu and said it was better than all of them.

The failures mattered too. Math was still limited: SAT-style story problems were consistent, college math and calculus appeared and then hit a ceiling, and the model was easy to confuse if Labenz posed as a student with a wrong answer. That made it a poor teacher, because a confused student could pull it off track. Chemistry equation balancing had the same pattern: basics worked, a planted student error often did not. Dates sometimes drifted in a Bing-like way. Across roughly a dozen roles, though, the cross-section ran from M.D. and lawyer down to hardware clerk and garage dispatcher, and Labenz's summary was that it did very well almost across the board. That reading, he notes, was later backed by two papers: *GPTs are GPTs*, where the second GPT means general-purpose technology, and a Microsoft Research paper arguing that GPT-4 is appropriately considered an early AGI.

The experiments are raw material. The economic inference is a three-step argument. First, if you look at jobs as they exist today and ask whether a language model can do the entirety of one, you will almost always conclude no, either because too much context is required or because some of the work is physical and the model has no robot. Second, that is exactly why Labenz expects a huge unbundling of jobs into tasks. The *GPTs are GPTs* paper takes the same lens: what tasks make up a role, how much time they take, and which of them a language model can do outright or cut at least in half. The finding is that about half of all jobs have about half of their tasks exposed.

Third, he reaches for manufacturing history. Before the Ford assembly line, people already made cars, but tolerances were loose and each vehicle needed artisanal runtime problem-solving. A modern semiconductor fab can be thrown off by a speck of dust. Knowledge work, Labenz argues, is about to get the same treatment: break the job down, see which parts a language model can do, train it until the output is reliable enough, and wire it into a system. Large companies will move first, because they will buy the most advanced systems and see the largest dollar savings; the pattern will then trickle down. He calls it a new Taylorism for knowledge work.

The arithmetic is the engine. If you pay people by the hour to take calls, a language model is at least a tenfold cost cut and often a 95 to 99 percent cut. A typical U.S. doctor's appointment at $100 — Labenz says that is probably conservative — can be a 45-minute conversation for about a dollar at then-current API prices. Call centers might be closer to ten-to-one than a hundred-to-one, but many things will land in that 10-to-100 range, plus 24/7 access, plus memory that does not forget who you are. Escalation is the first remaining human slot: "can I talk to a manager" is where first-generation models hand off. Labenz guesses that starts around 5 percent of cases, then 1 percent, then 0.1 percent.

A clinic receptionist is his concrete picture. Today the same person may answer the phone, schedule, and then take weight and blood pressure at check-in. He thinks it will be months at most before there is a competitive market in automated scheduling you can call anytime. The patient gets better service, the office saves money, and in theory some of that is passed through. Whether a job remains for the person who used to do all three depends on scale: in a small office the doctor may pick up vitals; in a large one three receptionists become one. Entrepreneurs, he says, are very good at filling those gaps, and they will do it faster than people expect. The fallacy is all-or-nothing thinking. It is not copilot versus replacement. Both happen at once.

Torenberg offers the classic consolation. Every technological wave was supposed to end jobs and then created more of them, because human desires are infinite. The United States went from something like 90 percent farmers to about 3 percent, and work did not disappear. He cites a Twitter exchange Labenz had with Antonio Garcia Martinez on that farmer chart, and asks whether the paradigm is unhelpful this time. Labenz says it is a tough question, he expects to be surprised, and he still thinks this time is different, because it is not clear what everyone is supposed to shift *into*.

He borrows a frame from Yuval Noah Harari: humans do physical labor with muscles and cognitive labor with brains, with some overlap and no obvious third place. Machines and fossil fuels took over much of the first, and people moved into the second. If cognitive work is now being taken over too, candidates for a third category are thin. Emotional labor is one; Labenz's own experiments suggest models are already pretty good at it, including the conversational pattern of cognitive behavioral therapy. A local, bespoke service economy — dinner-party murder mysteries, highly individualized entertainment — is another. An AI can design the mystery; people may still prefer to play it with other people. That is a hope, not a labor-market plan.

He offers a working definition of work: things people do not necessarily want to do, that therefore have to be paid for, that will not get done otherwise, and that only humans can do. That category, he says, is shrinking. He is not worried about his own boredom. If everything he does were automated he would still study AI and read the history books he currently has no time for. Anyone who ever wanted piano lessons and felt too busy might finally have the time, and the AI might give the lesson. The sting is the next clause: there is not necessarily money waiting at the end of that performance. A rich life and a paycheck are different questions.

Torenberg worries about a Harry Potter split into "Muggles" and a supercharged class, and recalls a conversation with Replit's Amjad Masad about the 1,000x developer. Labenz agrees software demand will rise, especially at very low prices, and then asks what software demand is actually infinite. Games and personalized entertainment are the honest end state, and even those collide with a 24-hour day. Original science and hardcore engineering are not in this generation's reach — a language model will not stand up its own semiconductor supply chain — but those rare specialists are already few, and the untouched category keeps shrinking. The speed is the other difference. Labenz is from Detroit. The textbook said the economy would adjust when auto jobs left; many Midwest towns never recovered, and the resentment became a political force. That was the slow version. The next two to five years, he thinks, compress that shock across many sectors at once. Societies have historically adapted via generational replacement. He does not know how you adjust if a huge share of employed people is hit in a few years.

Regulation, Labenz says, will not stop the transformation, but it will shape who can do what, under whose license, and on what timeline. He has been surprised by how slow interest groups have been; denial has outrun action. He does not expect that to last. Leadership at the AMA and the ABA will look at data like this and try to protect members. Medicine and law without a license are already illegal; the next move is to make unlicensed medical or legal language models illegal, or at least to require a licensed supervisor. The FDA has already made some noise about regulating language models as devices, which means a safety showing. Labenz would be shocked to get through the year without the first fights over who may use these models, who must supervise, and who is liable if OpenAI's product dispenses medical advice.

His rule of thumb is blunt: the higher a profession's current status, the more effective it will be at creating restrictions. Customer service is not organized that way. The McDonald's drive-through will likely be language-model powered, and nobody in Congress will take it up. If restrictions do not bite, he expects leapfrogging in places with few doctors per capita — he names the Democratic Republic of the Congo — where people have phones and questions, and the tools then work their way up the income ladder as it becomes obvious that blocking them is not in the public interest. Torenberg is more cynical. He thinks panic will scale, education will keep prioritizing teachers over students under the language of equity, and doctors, teachers' unions, and lawyers have enough power to do more harm than one would want. He does not think they can materially slow OpenAI. He thinks you get Marc Andreessen's chart: heavily regulated sectors (housing, healthcare, education) keep getting more expensive while unregulated ones get cheaper, and the fork widens.

Labenz grants the short-run fight and doubts a permanent lock. Nuclear plants and the Second Avenue subway can be stopped because they sit in one place. A recent EconTalk episode on state formation, he says, contrasted grain regions with tuber regions: grain can be stored and taxed, potatoes rot, so early states formed where you could tax. AI is a tuber. It is not in one location. China has a Great Firewall and still has VPNs. If the United States banned consumer medical models, OpenAI could run a Nairobi subsidiary and people would still reach it unless the regime became extremely heavy-handed. Demand will punch through the cracks. In the middle of the night, in winter, with a sick child, a conversation at one percent of the cost is something people will be motivated to find. The endgame, he thinks, is speed, not whether an AI doctor can ultimately be consulted.

Capability also outran the story. Labenz describes a counterfactual in which models improved slowly enough for the public to live with a lousy AI doctor and keep it off the market. Instead the last year crashed through that world. A paper that morning had GPT-4 beating Med-PaLM; Med-PaLM 2, out in the previous week or two, was neck and neck, both at expert level on standardized medical questions. People are hearing about this for the first time when it is already expert-level, which makes incumbent power grabs easier to see as naked. Fear, uncertainty, and doubt could still scare consumers for a few years. He would not be stunned to be arguing in 2028. Societies have jailed people for non-violent crimes on high-level safety arguments, and then mostly backed off the war on drugs. On AI doctors, he thinks the question of who is really being harmed and who is being protected will get even more obvious. His optimistic note is that five years is already a long time to hold the line against direct access.

Torenberg turns to morale. For a couple of decades the anxiety was blue-collar automation and what truckers would do for meaning, including debates over UBI. It turns out those jobs lasted longer than expected. White-collar work is more exposed instead, and the people in it are more sensitive to usefulness and status. They spent a lifetime climbing ladders and being told they were successful, special, and important. Torenberg admits he has felt a little deflated watching ChatGPT dwarf his own ability to synthesize, summarize, and analyze. The question is less about resources than about how people will feel needed after going through school, grad school, and the rest of the hoops.

Labenz treats that as another leveling force. A political strain of resentment has long targeted people who claim they got there on merit. Soon, he says, more of us will be in a similar boat, which might even make people a little nicer to each other. The reactions will still split. He had just seen a 3D artist — Torenberg had sent it — who lost everything he loved about the job and now sits prompting Midjourney all day, feeling less creative. The other half of the trade is real too: people who always had stories in their heads and could never visualize them, and now can. Both sides exist.

He is also watching a philosophical flip. Humans once distinguished themselves from animals by intellect and treated that as a license to run the world. The reverse claim is now that what makes us special relative to AIs is feeling, emotion, the animal traits we are pretty sure the models do not have. The awkwardness is that those traits are largely shared with animals. What separates us from animals puts us with the AIs; what separates us from the AIs puts us with the animals. For the moment humans sit in the center of that Venn diagram, and a loss of specialness around intellect is going to be a long psychological weather system.

A viral TikTok made the two truths sit in one frame. A doctor sat with GPT-4, watched it spit out a strong diagnosis, and crumpled: he went to medical school for four years and the thing just produces it. The video exploded. The top comments Labenz saw, all from women, were some version of "maybe at least ChatGPT will listen to me when I go in and talk about my problems." The doctor's status is in jeopardy; patients have a pent-up sense that they were not served well and that a more patient listener would be an upgrade. Both are true on the same screen. Labenz's guess about where life satisfaction goes is relationships, which every wellness study treats as a major variable, commercialized a little as highly local, idiosyncratic services — as long as something like Replika does not get so good that it crowds that out too.

Labenz does not claim to know what happens to society. He does think everyone is being pushed toward a new equilibrium in which you cannot really opt out. If every other doctor's office in an economy offers 24/7 scheduling, how do you remain the practice that takes calls from nine to four, charges more, and sells that as a reason to stay? The incentives all point toward rapid adoption. Feelings about status and meaning, he says, will have to be sorted out separately from how the work actually gets restructured.

Torenberg closes on a political-economy note that even people as left-wing as Ezra Klein have been promoting: well-intentioned liberals have made government so complex that it becomes a vetocracy, easy to veto and hard to build. The next few years will test how much power agencies and professional organizations actually have when a better deal for the customer is obvious. Will they prioritize the constituent, or the provider, even when the cost to the public is clear?

Labenz's note to listeners is that this is an honest take on change nobody is ready for. He quotes Tyler Cowen: a lot of copes are coming from a lot of directions, and this show is trying to be a cope-free zone. The point is to understand the technology on its own terms, what it can and cannot do, and how it is likely to be applied given those strengths and remaining weaknesses. That project, at this moment, does not yield tidy answers or a comfortable high-confidence outlook. It leaves the audience with the work of thinking it through in an environment that is still full of uncertainty.