# Dwarkesh's guests on work disappearing Research note for Jacob Steeves' Exploit Summit keynote (Montreal, 28 Sep 2026). Written 24 Sep 2026. Dwarkesh Patel hosts the Dwarkesh Podcast (dwarkesh.com). He has spent 2025 and 2026 arguing that the nature of work is about to change very fast, and that a large share of human jobs will stop being needed. The sharpest versions of that claim are not his. They come from the people he interviews, and from essays those people wrote on their own. Every line inside quotation marks below was read from the linked page on 24 Sep 2026. A line marked "reported" was carried by news outlets; the original page was not opened. Anything not checked is marked "unverified". Words used below, once: - **White-collar work**: jobs done at a desk, with language, numbers, and a computer. Law, finance, consulting, software, analysis. - **Labor share**: the slice of national income paid as wages. The rest is paid to owners of capital (machines, chips, shares, data centers). - **Drop-in worker**: an AI you can hire the way you hire a person. It uses Slack, email, and the company software, and finishes a project without you sitting next to it. - **Continual learning**: the ability to get better at a job over weeks, the way a new hire does, and to keep that learning. Today's models mostly forget when the chat ends. - **AGI**: in this note, an AI that can do almost any remote job at least as well as a skilled person. Speakers do not all use the word the same way. The numbers already in [the economics brief](/cursor/stores/bc-96ff1b3e-00b6-4da6-8a0d-5e72965f7e86/docs/research/economics-of-ai-and-labor-power.md) (labor share, young-worker hiring, firm valuations) are not repeated here. This note is the people and their own words. --- ## Executive summary 1. The claim Jacob wants is said most clearly by **Dario Amodei** (Anthropic), **Sholto Douglas** and **Trenton Bricken** (Anthropic researchers), **Leopold Aschenbrenner**, and **Philip Trammell** with Dwarkesh himself. They do not all mean the same thing, and they do not all mean "this year." 2. Amodei's own line, restated in his January 2026 essay: AI "could displace half of all entry-level white collar jobs in the next 1–5 years." He adds that AI is not a tool for one job. It is "a general labor substitute for humans." 3. Douglas, on the 22 May 2025 episode: a "drop-in white collar worker" is "very likely in two" years and "almost overdetermined in five." Bricken: even if progress stops, white-collar tasks "should" be automated "within the next five years," because the data is collectable and the wages are huge. 4. Aschenbrenner, June 2024, in *Situational Awareness*: the product is not a better chatbot. It is "a drop-in remote worker." He expects a "sonic boom": the jump in economic value arrives in a lump, because firms will not have finished integrating the weaker models before the strong one shows up. 5. Trammell and Patel, 29 Dec 2025, *Capital in the 22nd Century*: once machines can do the work, the old brake on inequality breaks. "Without [a global tax on capital], once AI renders capital a true substitute for labor, approximately everything will eventually belong to those who are wealthiest when the transition occurs, or their heirs." 6. Dwarkesh agrees with the end state and disagrees with the speed. His own view, June 2025: if progress stopped today, "less than 25% of white collar employment goes away," because models cannot learn a specific firm's habits. The bottleneck he names is continual learning. When that is solved, he thinks one model can learn every white-collar job at once, and firms become copies of a single mind. 7. On the same show, three guests push the other way. **Tyler Cowen** (9 Jan 2025): humans are the bottleneck, and explosive growth is unlikely. **Andrej Karpathy** (17 Oct 2025): the episode is titled "AGI is still a decade away." **Ege Erdil** and **Tamay Besiroglu** (17 Apr 2025): the episode is titled "AGI is still 30 years away." Their research still says that *if* full automation arrives, growth explodes and the labor share falls. 8. What has actually happened, by Anthropic's own paper of 5 Mar 2026: no rise in unemployment in the most exposed jobs since late 2022. Hiring of workers aged 22–25 into those jobs is down about 14% versus 2022, and that estimate is barely statistically significant. Amodei himself says displacement is not what is happening right now. He says it is what happens in the next one to five years. 9. Safe to say from the stage: the people building the systems expect desk work to be replaceable inside a few years, and they expect the income to go to owners of the machines. Do not say the labor market has already broken. Their own data says it has not. 10. The line that matches this talk: Amodei, January 2026. "Democracy is ultimately backstopped by the idea that the population as a whole is necessary for the operation of the economy. If that economic leverage goes away, then the implicit social contract of democracy may stop working." --- ## The episodes that are actually about this Not every Dwarkesh episode touches work. These do. Dates are the podcast's own. | Date | Episode | Who | What they add | |---|---|---|---| | 14 and 26 Jun 2023 | Carl Shulman, parts 1 and 2 | Carl Shulman | The early, full version: an intelligence explosion, then robot doublings. Human labor stops being the limit. Transcripts: [part 1](https://www.dwarkesh.com/p/carl-shulman), [part 2](https://www.dwarkesh.com/p/carl-shulman-2). No fresh verbatim labor line pulled in this pass. | | 4 Jun 2024 | Leopold Aschenbrenner — 2027 AGI | Leopold Aschenbrenner | Drop-in remote worker. His essay is the source; the episode is the walkthrough. [Episode](https://www.dwarkesh.com/p/leopold-aschenbrenner). | | 28 Mar 2024, 25 Mar 2025, 22 May 2025 | Sholto Douglas and Trenton Bricken (three episodes) | Both, Anthropic | The May 2025 episode has a section titled "Automating white collar work" at 2:10:26. [May episode](https://www.dwarkesh.com/p/sholto-trenton-2). | | 19 Feb 2025 and 12 Nov 2025 | Satya Nadella (two episodes) | Satya Nadella, Microsoft | How Microsoft is preparing. His public line is that work changes and people stay in the loop. He is not in the "people are obsolete" camp. Do not quote him as if he were. | | 3 Apr 2025 | AI 2027 | Scott Alexander, Daniel Kokotajlo | A written scenario, not a measurement. Late 2026: junior software jobs "in turmoil." [Scenario](https://ai-2027.com/). [Episode](https://www.dwarkesh.com/p/scott-and-daniel). | | 29 Apr 2025 | Mark Zuckerberg — AI will write most Meta code in 18 months | Mark Zuckerberg | The episode title is the claim. A verbatim sentence from the transcript was not pulled. [Episode](https://www.dwarkesh.com/p/mark-zuckerberg-2). | | 1 May 2025 | What will automated firms look like? | Dwarkesh Patel (narration of his own essay) | Firms with no human employees. Copies of one CEO replace the management layer. [Essay](https://www.dwarkesh.com/p/ai-firm). | | 17 Apr 2025 | Ege Erdil and Tamay Besiroglu — AGI is still 30 years away | Epoch AI | The slow timeline. Their papers still model what full automation does to growth and wages. | | 9 Jan 2025 | Tyler Cowen — the number one bottleneck is humans | Tyler Cowen | The main economist on the show who does not expect a fast wipeout. | | 12 Sep 2025 | Sergey Levine | Sergey Levine, UC Berkeley | Robots. His line is human-plus-robot first, not a switch that fires everyone. | | 17 Oct 2025 | Andrej Karpathy — AGI is still a decade away | Andrej Karpathy | The technical skeptic on timing. | | 13 Feb 2026 | Dario Amodei — "We are near the end of the exponential" | Dario Amodei | White-collar work, on-the-job learning, and why revenue is still tiny next to the wage bill. [YouTube](https://www.youtube.com/watch?v=n1E9IZfvGMA). | | 5 Feb 2026 | Elon Musk — cheapest place for AI is space | Elon Musk | This episode is about compute in orbit, not jobs. His "work will be optional" line is from a different event (below). | | 4 Jun 2026 | Alex Imas and Phil Trammell — What remains scarce after AGI? | Alex Imas (Chicago Booth), Philip Trammell (Stanford) | Wages, the labor share, and what is still scarce when machines do the work. [Episode](https://www.dwarkesh.com/p/alex-imas-phil-trammell). | Dwarkesh's own monologues on the same subject, which are not interviews: - [Why I don't think AGI is right around the corner](https://www.dwarkesh.com/p/why-i-dont-think-agi-is-right-around) (2 Jun 2025, narrated 3 Jul 2025). Continual learning is the bottleneck. - [Timelines, June 2025](https://www.dwarkesh.com/p/timelines-june-2025). He argues with Douglas and Bricken by name. - [Thoughts on AI progress (Dec 2025)](https://www.dwarkesh.com/p/thoughts-on-ai-progress-dec-2025). - [The next big breakthrough will be AIs learning on the job](https://www.dwarkesh.com/p/the-next-paradigm) (26 Jun 2026). --- ## What those people wrote themselves ### Dario Amodei — the plainest warning CEO of Anthropic. Two Dwarkesh interviews: 8 Aug 2023 and 13 Feb 2026. The labor claim lives in his essays, not in the 2023 interview. **Axios interview, 28 May 2025.** Reported by Mike Allen and Jim VandeHei, and carried across the press the same day. The original Axios page timed out on retrieval, so treat the numbers as reported, not as a line read from Axios: AI could wipe out half of entry-level white-collar jobs, and unemployment could go to 10–20%, within one to five years. Danger zones he named: tech, finance, law, consulting. He said companies and government should stop "sugar-coating" it. **His own restatement.** *The Adolescence of Technology*, January 2026, published 26 Jan 2026. [darioamodei.com/essay/the-adolescence-of-technology](https://darioamodei.com/essay/the-adolescence-of-technology). He defines the system he is worried about as "a country of geniuses in a datacenter": smarter than a Nobel winner across most fields, able to use a computer the way a person does, able to take a task that lasts weeks, and cheap enough to run in millions of copies. On jobs, in his words: > "I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress." > "AI isn't a substitute for specific human jobs but rather a general labor substitute for humans." Why he thinks this is not the old story of farmers becoming factory workers: - **Speed.** In two years, models went from barely writing a line of code to writing almost all the code for some Anthropic engineers. People and labor markets move slower than that. - **Breadth.** Finance, consulting, and law use the same kind of mind. If only one of them is hit, people switch. If all three are hit at once, they cannot. - **Who gets hit.** The models climb from weak work to strong work. He fears an "unemployed or very-low-wage underclass" sorted by cognitive ability, which is harder to retrain than a job title. - **Gaps close.** A weakness that used to protect a human task (he uses image models drawing the wrong number of fingers) now gets fixed in months, because the labs train on the gap. He grants slow adoption inside old firms. That is why his number is one to five years, not this quarter. He does not think slowness saves the end state. Startups can route around the old firms. He calls one result "geographic inequality": wealth piles up where the models are built. On comparative advantage (the textbook point that even a weaker party still has something to trade): > "The problem is that if AIs are literally thousands of times more productive than humans, this logic starts to break down." On power, which is the sentence for this talk: > "Democracy is ultimately backstopped by the idea that the population as a whole is necessary for the operation of the economy. If that economic leverage goes away, then the implicit social contract of democracy may stop working." > "The thing to worry about is a level of wealth concentration that will break society." He says Elon Musk's fortune, about $700 billion at the time of writing, already exceeds John D. Rockefeller's peak as a share of the economy. Rockefeller was about 2% of US GDP. He is explicit that this is before most of the AI money arrives. He also says the end state is total: "Ultimately, I think of all of the above interventions as ways to buy time. In the end AI will be able to do everything." His proposed buys of time: measure displacement as it happens (Anthropic's Economic Index), steer customers toward doing more rather than employing fewer, look after his own staff even after they "are no longer providing economic value in the traditional sense," and a progressive tax. He says if the wealthy refuse a well-designed tax, they will get a badly designed one "designed by a mob." **Check against his own lab.** Anthropic, *Labor market impacts of AI*, 5 Mar 2026. [anthropic.com/research/labor-market-impacts](https://www.anthropic.com/research/labor-market-impacts). No systematic rise in unemployment for the most exposed jobs since late 2022. For workers aged 22–25, the rate of finding a new job in the most exposed occupations is about 14% lower than in 2022. The authors say that estimate is barely statistically significant. There is no such drop for workers over 25. Most exposed occupations they name: computer programmers, customer service representatives, financial analysts. Amodei agrees in the essay that displacement is likely not happening yet. The essay is a forecast. The paper is a measurement. Use both. ### Sholto Douglas and Trenton Bricken — the five-year drop-in worker Researchers at Anthropic. Douglas works on scaling reinforcement learning. Bricken works on interpretability (reading what a model is doing inside). Their own public writing on this is the podcast. There is no separate essay. From the 22 May 2025 episode, [dwarkesh.com/p/sholto-trenton-2](https://www.dwarkesh.com/p/sholto-trenton-2). Section "Automating white collar work," timestamp 2:10:26. Douglas: > "The one that I feel we're almost guaranteed to get — this is a strong statement to make — is one where at the very least, you get a drop-in white collar worker at some point in the next five years. I think it's very likely in two, but it seems almost overdetermined in five." > "Plan for the case where white collar work is automateable." > "We're in for a much worse world if it just so happens that the people who had money in the stock exchange, or in land before AGI are dramatically more wealthy than the people who don't. It's a gross misallocation of resources." His policy point for a country that does not have a frontier lab: the valuable thing becomes compute (the chips you run the model on), and you want rules that stop capital from locking in before everyone else can buy any. Bricken, same episode. Dwarkesh later quotes this back in his June 2025 timelines post, and the transcript matches: > "Even if AI progress totally stalls, you think that the models are really spiky, and they don't have general intelligence. It's so economically valuable, and sufficiently easy to collect data on all of these different jobs, these white collar job tasks, such that to Sholto's point we should expect to see them automated within the next five years." Bricken's mechanism is not "the model becomes generally intelligent." It is "the wages are so large that someone will record the tasks and train on them." He also flags Moravec's paradox: the things people think are impressive (desk work) may fall first, and the things people do not think about (hands, walking, care) may remain. He thinks that paradox is "a little bit fake," because desk work has GitHub and the internet, and physical work does not yet have the same recordings. If those recordings existed, he thinks robots would move at a similar speed. On the 25 Mar 2025 AMA they give career advice under this forecast: learn by using the models, and expect one person's job to become "managing" many copies. They do not name a safe profession. Douglas jokes that podcasting might last. That is a joke, not a finding. ### Leopold Aschenbrenner — the sonic boom Former OpenAI researcher. Essay series *Situational Awareness*, June 2024. [situational-awareness.ai](https://situational-awareness.ai/). Dwarkesh interview the same month: [dwarkesh.com/p/leopold-aschenbrenner](https://www.dwarkesh.com/p/leopold-aschenbrenner). The labor claim, from the essay PDF: > "By the end of this, I expect us to get something that looks a lot like a drop-in remote worker. An agent that joins your company, is onboarded like a new human hire, messages you and colleagues on Slack and uses your softwares, makes pull requests, and that, given big projects, can do the model-equivalent of a human going away for weeks to independently complete the project." > "The drop-in remote worker will be dramatically easier to integrate — just, well, drop them in to automate all the jobs that could be done remotely. It seems plausible that the schlep will take longer than the unhobbling ... so the jump in economic value generated could be somewhat discontinuous." **Schlep** here means the dull work of changing a company's habits. **Unhobbling** means removing the limits that keep a strong model stuck as a chatbot (no memory, no computer use, no long tasks). His point: firms will still be wrestling with clumsy tools when a hireable agent arrives, so the economic break can look sudden even if the research was smooth. He dates the underlying capability jump to about 2027. That date is a forecast, and it is the one Karpathy, Erdil, and Besiroglu reject. ### Philip Trammell and Alex Imas — what the income does Trammell is an economist at Stanford. Imas is an economist at Chicago Booth. They are the guests who do this as economics rather than as a lab timeline. **The paper, with Anton Korinek.** *Economic Growth under Transformative AI*, NBER working paper 31815, October 2023, revised through 2026. [nber.org/papers/w31815](https://www.nber.org/papers/w31815). The abstract's result, in their words: fully automating production, so that machines can copy themselves, "would dramatically raise the growth rate and lower the labor share." Wages might rise or fall. Output can explode while the share paid to people shrinks. This is the same mechanism already summarized in the economics brief (Korinek and Suh: wages can collapse even as output grows more than tenfold). Trammell is the person on the podcast who co-wrote it. **The essay, with Dwarkesh.** *Capital in the 22nd Century*, 29 Dec 2025. [philiptrammell.substack.com/p/capital-in-the-22nd-century](https://philiptrammell.substack.com/p/capital-in-the-22nd-century). Their claim about the past: Thomas Piketty was wrong that capital accumulation by itself drives inequality forever. For centuries, capital and labor needed each other. More machines made workers more valuable. That pulled wages up and stopped the rich from pulling away forever. They call capital and labor **gross complements** in that world. Their claim about the future: advanced AI makes capital a substitute for labor. Then the brake fails. Piketty "will probably be right about the future." > "Without one, once AI renders capital a true substitute for labor, approximately everything will eventually belong to those who are wealthiest when the transition occurs, or their heirs." The "one" is "a global and highly progressive tax on capital (or at least capital income)." Two details that matter for a room of investors: - A house is a bad asset in that world. Most of a normal household's wealth is a house. A house does not train models or run robots. Public stocks are already very unequally held (they cite a Gini above 0.9, and 42% of Americans with no stocks in 2022). Private AI startups are held more narrowly still. You cannot buy xAI in a 401(k). The Sultan of Oman can. - Developing countries lose their usual catch-up path. That path was cheap labor plus imported machines. If labor is not the scarce input, cheap labor is not an advantage. **The episode, 4 Jun 2026.** [dwarkesh.com/p/alex-imas-phil-trammell](https://www.dwarkesh.com/p/alex-imas-phil-trammell). They are more careful live than the essay's last sentence. Trammell says nothing has been fully automated yet, once you count the whole supply chain. Computer and electronics in the US still have a capital share around half, not 100%. The qualitative break is a good whose entire supply chain needs no human, and that nobody insists a human do for its own sake. Imas's caution, same episode: automating nine of ten tasks can make the remaining human task more valuable, if demand rises enough. He also describes a messy middle, using phone operators from 1920 to 1940. The job vanished over twenty years, people were rehired, and they were rehired at lower pay. He says a fast spike in unemployment of even 2–3% becomes a political emergency. A slow slide into worse jobs may not. Imas is the person on this show who says the bad outcome can look like underemployment, not a headline unemployment number. ### Dwarkesh Patel — agrees on the destination, not the year His essay *What fully automated firms will look like* (31 Jan 2025; narrated 1 May 2025). [dwarkesh.com/p/ai-firm](https://www.dwarkesh.com/p/ai-firm). The picture: you do not hire a thousand AI workers of mixed quality. You take one excellent mind and copy it. A human CEO cannot watch 200,000 people. An AI copy of that CEO can, if you give it the chips. Middle management is replaced by copies. The scarce input is compute, not talent. He thinks firms get larger, because talking to an exact copy of yourself is cheaper than hiring. His brake, stated in June 2025 against Douglas and Bricken. [dwarkesh.com/p/timelines-june-2025](https://www.dwarkesh.com/p/timelines-june-2025): > "If AI progress totally stalls today, I think <25% of white collar employment goes away." Reason: a model that cannot learn his preferences still loses to a human editor. He will not call something an employee until it can be on the job for months and get better. His date for that, in the same post: an AI video editor that after six months knows his show as well as a person would, in **2032**. So the honest summary of the host: the wipeout is real once models learn on the job. It is not real from today's models plus more task recordings. That is a direct contradiction of Bricken, said on purpose. On the 13 Feb 2026 Amodei episode he puts a scale on it. Human wages are on the order of $50 trillion a year. Lab revenue is a rounding error next to that. His point to Amodei: if the models could already do the jobs, the revenue would be there. ### The guests who say not so fast Use these so the claim does not sound like the whole show agrees. - **Tyler Cowen**, 9 Jan 2025, [dwarkesh.com/p/tyler-cowen-4](https://www.dwarkesh.com/p/tyler-cowen-4). Humans are the bottleneck. Diffusion through real firms is slow. He does not expect AI to produce explosive economic growth. He has said versions of this on the show since 2024. - **Andrej Karpathy**, 17 Oct 2025. Episode title: "AGI is still a decade away." His June 2025 talk *Software Is Changing (Again)* is already in the store. That talk is about programmers rewriting software, not about the wage bill going to zero. - **Ege Erdil and Tamay Besiroglu**, 17 Apr 2025. Episode title: "AGI is still 30 years away." They run Epoch AI. Their written work still says that full automation, if it arrives, breaks the old growth regime. The disagreement is the clock, not the mechanism. - **Sergey Levine**, 12 Sep 2025, [dwarkesh.com/p/sergey-levine](https://www.dwarkesh.com/p/sergey-levine). Physical work. He expects robot-plus-human for a long time, the way coding assistants raised output without firing every engineer at once. Scope grows. It does not flip. - **Satya Nadella**, 19 Feb 2025 and 12 Nov 2025. Two episodes on Microsoft preparing for AGI. His public product line is a copilot sitting next to a worker. Do not draft him into Amodei's sentence. ### Adjacent, and easy to mis-cite - **Scott Alexander and Daniel Kokotajlo, AI 2027** (published April 2025; episode 3 Apr 2025). A scenario they wrote so people could argue with a concrete story. In that story, late 2026 brings junior software jobs "in turmoil," and by mid-2027 startups promise to disrupt every white-collar profession. They also write a branch where the build-out is slowed. Quote it as their scenario, not as data. - **Elon Musk.** On the Dwarkesh episode of 5 Feb 2026 he talks about putting computers in space. The jobs line is from the U.S.–Saudi Investment Forum in November 2025, reported by Fortune and others: "My prediction is that work will be optional." He compares a job to growing vegetables when you could buy them. That is a different event. Do not attribute it to the podcast. - **Mark Zuckerberg**, 29 Apr 2025. The episode is titled "AI will write most Meta code in 18 months." That is a claim about one company's code, not about the labor share. An earlier episode (18 Apr 2024) is about Llama 3 and data centers. A checked verbatim line was not pulled. - **Carl Shulman**, June 2023. This is the episode Dwarkesh still points people to for the full "intelligence explosion, then the economy is robots" argument. It is two long interviews. A fresh transcript quote was not extracted here. If a line of his is wanted on a slide, pull it from the episode page before using it. - **Demis Hassabis** (28 Feb 2024), **Ilya Sutskever** (27 Mar 2023 and 25 Nov 2025), **Shane Legg** (26 Oct 2023). They are on the show about whether the systems get built. They are not the labor argument. --- ## What is safe to say in Montreal The room already knows AI. The useful move is to separate three claims these people run together. **Claim A. The work changes fast.** Douglas, Bricken, Aschenbrenner, and Amodei. Their window is about two to five years for desk work, once an agent can finish a job rather than a prompt. Karpathy says a decade. Erdil and Besiroglu say thirty years. Dwarkesh says the switch is continual learning, and his own date for a true AI employee is 2032. **Claim B. The people become unnecessary.** Amodei's phrase is "general labor substitute." Trammell and Korinek's result is a labor share that falls, possibly toward zero, while output rises. Imas says the visible form may be worse jobs, not a unemployment spike. Anthropic's March 2026 paper says the spike has not happened. **Claim C. The power moves to whoever owns the machines.** This is the claim that matches the rest of the talk, and it is the one these speakers agree on even when they fight about the year. Trammell and Patel: the heirs of today's capital owners. Douglas: people who already hold stocks and land. Amodei: a fortune can already match the Gilded Age, and democracy assumed that citizens were needed for the economy. Dwarkesh's firm essay: the scarce input is compute, and one mind can be copied across the whole company. A is a timing fight. B is not yet in the unemployment rate. C is the point. Lines that will be quoted back at you if you overshoot: - "Half of entry-level white-collar jobs" is Amodei's forecast for one to five years from May 2025. It is not a current statistic. - "10–20% unemployment" is reported from the Axios interview. It is not in the January 2026 essay passages read here. If you use it, say he told Axios. - "Work will be optional" is Musk in Riyadh, not Musk on this podcast. - Dwarkesh will not back "most white-collar jobs are already gone." He wrote the opposite. One sentence that is his, checked, and on the talk's subject: > "If that economic leverage goes away, then the implicit social contract of democracy may stop working." Dario Amodei, *The Adolescence of Technology*, January 2026.