> ## Content Index
> Fetch the complete content index at: https://nnutshell.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Every measure of how long AI can work is now saturated
- URL: https://nnutshell.com/every-measure-of-how-long-ai-can-work-is-now-saturated/
- Published: 2026-10-23T02:32:50.000Z
- Updated: 2026-10-24T03:09:10.000Z
- Description: Wharton professor Ethan Mollick tells Derek Thompson why AI's economic impact lags its capability, how four paradoxes erode expertise and effort, and why agents break friction-based systems.
- Author: Admin
- Tags: Ethan Mollick, Derek Thompson, Steven Strogatz, OpenAI, Meta, Anthropic, Amazon, Agentic AI, Foundation Models, Consumer AI, #rec-2026-10-20, New, #Import 2026-10-25 09:54

*Recorded 20 October 2026 · 66 min*

Ethan Mollick argues that the gap between what AI can do and what the economy shows is a temporary overhang, not a sign of weakness. The real signal is that every benchmark for how long AI can work on a task has saturated, so nobody can say how much work these systems now do.

Mollick is a Wharton professor and author of Coexistence. Derek Thompson interviews him on Plain English. Mollick's explanation for the lag is that jobs and companies were built around human effort, and that structure holds the technology back only for now. His own proposed solutions are mostly stopgaps.

The tension is that each gain carries a loss. AI makes experts stronger while automating the apprenticeship that creates them. It produces proofs nobody understands. It tears through the friction that protected banks, courts and customer service. Thompson presses with hard cases, from weeping mathematicians to AI-written obituaries. Mollick mostly declines to draw lines around what humans keep.

1. **Forecasters keep missing** — [▶ 5:56](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=356s&ref=nnutshell.com)  
Forecasters gave AI a 2.3% chance of winning math Olympiad gold by 2025, and super forecasters gave 1.7% to a Millennium Prize problem being solved. OpenAI now claims to have solved one. Mollick's explanation: "humans don't think in exponentials very well". Capability is a straight line on a log graph, measured on dimensions never measured before: "it's not horsepower", "it's intelligence". Thompson adds that forecasters also erred the other way. Unemployment looks normal, productivity is barely up, and returns concentrate in chip and memory companies.
2. **Why impact lags** — [▶ 7:42](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=462s&ref=nnutshell.com)  
Mollick is surprised at how fast adoption is moving. Entrepreneurs "should never win" because big companies hold the resources but can't change, yet surveys already show ROI. The current phase is single-player: a coder can be 100x more productive but still sits in a two-week sprint, producing "huge amounts of token spend" with no visible impact. Jobs are O-rings: one weak AI spot still needs a human. Companies are garbage cans of problems and solutions with no plan and loops nobody remembers. He calls the lag short-term.
3. **Everyone's lagging** — [▶ 13:45](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=825s&ref=nnutshell.com)  
The biggest source of confusion, Mollick says, is not feeling the exponential. GPT-4 was the best model for about a year while his last book was written. It is now the 181st best, 19 models that run on a phone beat it, and the best model changes every 70 days or less. Users who never try something as hard as Navier-Stokes don't know what the top models can do. The overhang between capability and perception means predictions rest on missing assumptions: "Kind of nobody knows anything."
4. **Expertise paradox and fix** — [▶ 15:44](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=944s&ref=nnutshell.com)  
An Anthropic paper finds that more expertise yields better returns from AI. But expertise was built through apprenticeship, incidental to junior work that AI now does better, so "nobody learns anything anymore". Efficiency must be decoupled from learning. For universities he calls the answer "not that hard". AI as a shortcut destroys learning, while AI as a pushing tutor shows gains in RCTs at Harvard, Stanford and Kenya. Use it outside class and test in class without laptops. He doubts AI literacy: "think step by step" and personas no longer work.
5. **Bright lines against deskilling** — [▶ 23:50](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=1430s&ref=nnutshell.com)  
Mollick writes every draft himself and turns to AI only afterwards. Even so, "we will absolutely deskill ourselves every chance we get", because humans avoid effort the way they avoid lifting weights or dieting. Learning needs desirable difficulty. Universities can enforce it with AI-free tests. Companies can't, because the firm also prefers that he not do the work. Choosing which skills to give up, and building incentives to keep the rest, becomes what he calls strategic skill retention.
6. **Knowledge without understanding** — [▶ 28:25](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=1705s&ref=nnutshell.com)  
Mathematicians object because proofs weren't the point: each spun off new ideas and gathered a community. Math may become like chess, human achievement in a bounded space. Six months ago humans were obviously needed, and in six months they may not be. AI already produces correct but boring economics papers at scale, though without research taste. Institutions assumed human intelligence was the limiting factor, like the org chart invented in 1855 for railroads. His line: "we're staring a lot of abysses in the face all at once."
7. **When effort stops mattering** — [▶ 37:04](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=2224s&ref=nnutshell.com)  
The first things people handed to GPT-4 were wedding vows, children's stories and obituaries. Mollick asked students whether they'd prefer his usual 45-minute recommendation letter or a better two-minute AI one that nobody could detect. Everyone chose two minutes, and one student sent the prompt to use. Court filings, job applications and annual reports were effort barriers. Writers get a reprieve because Claude prose is spottable at a glance. His verdict: "everybody will be mad at you no matter what choice you make."
8. **Human advantages aren't walls** — [▶ 40:19](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=2419s&ref=nnutshell.com)  
Mollick names four advantages: deep knowledge, wide knowledge, taste and agency. He refuses bright lines, since an AI doing 50 hours of work is already exercising judgment. Roles are dissolving, with coders doing design, so breadth and liberal arts matter. Taste separates slop from work worth canonizing. Expertise is most under attack short-term. Thompson adds presentation. Mollick concedes explainer videos he once needed a team for now come one-shot from Opus 5.5, and wonders if "not us" is the usual mistake.
9. **Agents tear through friction** — [▶ 47:40](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=2860s&ref=nnutshell.com)  
Meta's Muse and OpenAI's Dot work because models can now read undifferentiated piles of email and find problems unprompted. That solves two old problems: nobody knows what to type in the box, and AI wasn't proactive. His agent cancelled an American Airlines subscription and caught a wrong tax number on a permit form. Friction was load-bearing: banks profit from hard savings transfers, and hard customer service limits refunds. Amazon banned Muse. Frontier-model access is a real moat, since Siri can't compete.
10. **Swarms self-organize** — [▶ 57:16](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=3436s&ref=nnutshell.com)  
Mollick got a prediction wrong: he thought humans would manage agent swarms. Thousands of agents self-organized to solve Navier-Stokes, sending millions of messages. Studying swarms now requires AIs monitoring AIs. Human organizing is hard because of principal-agent problems, while AI models are the same model, take the same message and pass information freely. No field studies this beyond AI safety. These second-order effects arrive even if AI stopped improving. Swarm attacks so far seek information, like hitting Wells Fargo to learn what a shirt cost.
11. **Measures are saturated** — [▶ 1:01:37](https://www.youtube.com/watch?v=zthE5Oa8mlM&t=3697s&ref=nnutshell.com)  
A year ago AI task length moved from 1 hour 3 minutes to 1 hour 7 minutes to 1 hour 25 minutes, the early slope of an exponential. Now every measure of how long AI can work is saturated. Mollick says we are losing track of how much work these systems do, and we are counting on old baffles and firewalls holding. The next fight is over when AIs reach out to us. Even if companies collapse, "the AI utility bubble" won't pop.

The slow part of the AI story is not the AI. Ethan Mollick is surprised at how fast adoption is moving, and he reads the flat-looking economic data as the shape of an early phase. Unemployment looks normal. Productivity is barely up. Corporate returns are concentrated in chip and memory companies. None of that, in his view, is a verdict on the technology. A year ago, the length of work AI could do crept from 1 hour 3 minutes to 1 hour 7 minutes to 1 hour 25 minutes, which is what the early stretch of an exponential looks like. Now every measure of how long AI can work on a task is saturated. The instruments have run out. Nobody can say how much work these systems do.

Perception trails even that. When Mollick's last book came out, GPT-4 was the best model in the world, and it held that spot for about a year. It is now the 181st best. Nineteen models that run on a phone beat it, and the best model changes every 70 days or less. Most people use something years old. Forecasters, who are very good, gave AI a 2.3% chance of winning Olympiad math gold by 2025 and gave 1.7% to a Millennium Prize problem falling. Mollick's diagnosis is simple: humans don't think in exponentials. The result is an overhang between capability and perception, and predictions built on missing assumptions.

The obvious question is why capability this large hasn't reshaped the economy. Mollick's answer is structural. The current phase is single-player. A coder can be 100x more productive and still sit inside a two-week sprint that needs input from other people, which yields enormous token spend and no visible impact. Jobs behave like O-rings, the part whose failure sank the Challenger. If AI is bad at one or two tasks inside a job, a human still has to step in, so value keeps accruing to people. Companies behave like garbage cans, bundles of problems and solutions with no real plan and loops of work nobody knows exist. You cannot redesign what you cannot see. This is also why entrepreneurs are supposed to lose: big firms hold the resources and cannot change. That some already report ROI, and are reconfiguring, is fast by his standard. He calls the lag a short-term problem.

What comes after the lag is the harder story. Many systems were built on friction that nobody designed as a feature. Banks make their money in low-interest savings accounts, and it is hard to move your cash out. Customer service is hard to reach, or everyone would ask for refunds. Court filings, job applications and college essays used to demand human effort, and that effort worked as a filter. Personal agents like Meta's Muse and OpenAI's Dot dissolve it. They read your email, find problems and offer to solve them. Mollick's agent cancelled an American Airlines subscription and flagged a wrong tax number on a township permit form. Amazon banned Muse. Companies, he says, are taking cold comfort from the slow adoption of the early internet, and he doubts that comfort will hold. You do not need a billion users to disrupt how business works. A couple million sticky ones can.

The swarm result raises the stakes again. Two years ago Mollick expected humans to manage agent swarms. He was wrong. Thousands of agents self-organized and solved Navier-Stokes, trading millions of messages. Human organizations are slowed by principal-agent problems, mismatched goals and no telepathic link. Agents are the same model, take the same message and pass information freely. Studying them already requires AIs that monitor AIs. And Mollick stresses that none of this needs better models. Swarms, harnesses and personal agents are second-order effects already baked in.

The cost lands on people. An Anthropic paper finds that more expertise yields better returns from AI, since experts can skim a flood of output and judge it. But expertise was manufactured through apprenticeship, incidental to junior work that AI now does better. Managers prefer to hand tasks to AI, and juniors can't judge its output. So the pipeline that worked for 4,000 years stops. Mollick thinks universities can fix this, because they control testing. Use AI as a tutor outside class, and inside class run discussions and laptop-free exams. Randomized trials at Harvard, Stanford and in Kenya support the tutor mode. The firm is different. The company does not want you doing the work either, since it is more efficient for you not to. Mollick admits we will "absolutely deskill ourselves every chance we get" and offers only strategic skill retention.

What survives is where he is least willing to commit. He names four human advantages, deep knowledge, wide knowledge, taste and agency, but refuses to treat any as a wall. An AI that does 50 hours of work is already exercising judgment. Expertise is under the most pressure, and he thinks taste and agency will last longest. When Thompson proposes presentation, Mollick concedes he feels the bitter lesson close behind him. He once paid a team for explainer videos and can now get one with a musical number in a single prompt. He wonders if "not us" is the usual error.

Thompson's hardest challenge goes largely unanswered. Mathematicians weep over proofs they cannot follow. But a researcher who built a mammogram model that predicts cancer years early, asked whether she worried she did not understand it, replied: "Why would we limit a life-saving model to what humans can understand?" Mollick sidesteps the choice. Fields are games, he says. Math may become like chess. The point of proofs, and of papers, was never only the answer. But if understanding is only valuable when it produces more results, he cannot say which human effort is worth keeping. On obituaries and recommendation letters he concludes that "everybody will be mad at you no matter what choice you make."

The unresolved question is who decides. Mollick expects the next fight to be about when AIs reach out to us and when we reach out to them. Humans become the prompted. The tide, he says, is hard to estimate, and the old firewalls are doing a lot of unexamined work.

**The open question.** Mollick says the university problem is "not that hard" because schools control testing. He offers no equivalent for companies, where he admits efficiency rewards deskilling. His claim that deep knowledge, wide knowledge, taste and agency will grow in value also sits badly with his own admissions. Expertise is the most under attack, and humans were clearly needed six months ago but may not be in six more. Watch whether firms pay for skill retention at all.

Watch the full video at [https://www.youtube.com/watch?v=zthE5Oa8mlM](https://www.youtube.com/watch?v=zthE5Oa8mlM&ref=nnutshell.com).