---
title: 'The Experts Were Wrong About AI. Again.'
source: 'https://youtube.com/watch?v=xdwFncdu8t0'
video_id: 'xdwFncdu8t0'
date: 2026-11-11
duration_sec: 442
channel: 'Sabine Hossenfelder'
---

# The Experts Were Wrong About AI. Again.

> Source: [The Experts Were Wrong About AI. Again.](https://youtube.com/watch?v=xdwFncdu8t0)

## Summary

The video examines how experts have consistently underestimated AI's progress in cognitive tasks, citing specific examples like solving Millennium Prize Problems and achieving high scores on programming benchmarks. It contrasts this with overestimations in physical tasks like trucking automation, and concludes that software-only domains will see rapid advancement while physical changes will take decades.

### Key Points

- **Millennium Prize Problem solved early** [00:47] — Experts gave a 10% median probability for AI solving a Millennium Prize Problem by 2027; AI hit this benchmark in September 2026.
- **LiveCodeBench Pro score 4x higher than predicted** [01:21] — Experts predicted a rise to 14% on LiveCodeBench Pro by end of 2026; AI reached 53.8% by May.
- **AI task-completion horizon exceeded predictions** [01:52] — Experts predicted a 3.4-hour task-completion horizon by end of 2026; a new model reached 3 hours 6 minutes during the survey.
- **Math theorem proving arrived much sooner** [02:10] — In 2023, experts estimated 22 years for AI to prove publishable math theorems; now AI-written papers flood preprint servers.
- **Autonomous vulnerability exploitation achieved early** [02:49] — Cybersecurity experts said autonomous vulnerability exploitation was years away in 2023; AI achieved it the next year.
- **Math Olympiad gold achieved ahead of schedule** [03:02] — Experts predicted gold-level Math Olympiad skills by 2030; AI reached it in 2025.
- **Automation overestimated** [03:46] — WEF predicted 50% automation by 2025; actual rate was 22%.
- **Intelligence explosion prediction failed** [04:00] — Aschenbrenner's intelligence explosion prediction by 2027 failed due to physical and institutional hurdles.

## Transcript

A decade ago, most of us would have been stunned&nbsp; by the idea that a computer program can write&nbsp;&nbsp; an entire book, let alone a scientific paper.&nbsp; But come September 2026 and the headlines read&nbsp;&nbsp;
“Using AI, a professor wrote 200 papers&nbsp; this year” and it doesn’t even surprise&nbsp;&nbsp; anyone anymore. We have gotten so used to AI&nbsp; so quickly that it’s easy to forget just how&nbsp;&nbsp; much it’s outperformed expectations. At least&nbsp; some of them. Today I have a quick rundown of&nbsp;&nbsp;
these wrong predictions, and some thoughts&nbsp; on what we can learn from this. Make sure&nbsp;&nbsp; you stick around till the end. I've got a&nbsp; great offer from today's sponsor, Outskill.&nbsp; Last year, a group of AI researchers asked more&nbsp; than 200 experts how likely it was that artificial&nbsp;&nbsp;
intelligence would solve or substantially help&nbsp; solve one of the Millennium Prize Problems by&nbsp;&nbsp; the end of 2027. The median answer was 10%.&nbsp; AI hit this benchmark in September 2026.
The same group of experts was asked how well&nbsp; the best AI would perform during 2026 on an&nbsp;&nbsp; extremely difficult programming test called&nbsp; LiveCodeBench Pro. At the time, the best score&nbsp;&nbsp; was 7.7%. The median expert prediction was that&nbsp; this would rise to 14% by the end of 2026. By May,&nbsp;&nbsp;
AI had reached 53.8%. So the actual result&nbsp; was almost 4 times the experts’ prediction,&nbsp;&nbsp; In April and May 2026, experts were asked about&nbsp; something called the AI task-completion time&nbsp;&nbsp;
horizon. Roughly speaking, this measures how&nbsp; long a coding job would take a human expert&nbsp;&nbsp; when an AI can still complete it reliably.&nbsp; The experts predicted that by the end of 2026,&nbsp;&nbsp; AI would reach about 3.4 hours.&nbsp; While the survey was still running,&nbsp;&nbsp;
a new model had already&nbsp; reached 3 hours and 6 minutes. This isn’t a recent thing, it’s&nbsp; been going on for a few years now. In autumn 2023, more than 2000 AI researchers&nbsp; were asked how long various advances would&nbsp;&nbsp;
take. They estimated it’d be 22 years until AI&nbsp; could prove mathematical theorems publishable in&nbsp;&nbsp; leading mathematics journals. Today’s reality&nbsp; is that the maths and physics preprint server&nbsp;&nbsp;
has announced they are being flooded with&nbsp; papers that are, by admission of the author,&nbsp;&nbsp; AI-written and that even the authors do not&nbsp; understand. The new part isn’t that they don’t&nbsp;&nbsp; understand their paper, the new part is that&nbsp; they admit it. And the question is not any&nbsp;&nbsp;
longer whether the papers are publishable,&nbsp; but who is responsible for the content. In 2023 cybersecurity experts said that&nbsp; AI wouldn’t be able to autonomously find&nbsp;&nbsp;
and exploit vulnerabilities for years&nbsp; to come. AI hit the mark the next year.&nbsp;&nbsp; In 2022, experts estimated that AI&nbsp; would reach gold level skills in the&nbsp;&nbsp; international Mathematical Olympiad&nbsp; by 2030. AI reached it in 2025.
So you see the pattern there, experts have&nbsp; for many years underestimated AI progress.&nbsp;&nbsp; My personal speculation for why this is is that&nbsp; many of the so-called experts work in academia.&nbsp;&nbsp; They’re used to things taking forever. They&nbsp; don’t understand that market economies operate&nbsp;&nbsp;
at a completely different pace, especially when&nbsp; bolstered by hundreds of billions of dollars. But there’s a flipside. If you look at what&nbsp; experts predicted would happen to jobs,&nbsp;&nbsp; they were largely wrong. For example, one of&nbsp; the jobs that was supposed to be automated&nbsp;&nbsp;
first were truckers. This hasn’t even really&nbsp; begun to happen. In 2020, the World Economic&nbsp;&nbsp; Forum expected nearly half of work to be&nbsp; automated by 2025. In a follow-up from 2025,&nbsp;&nbsp;
they found that the actual rate was more like 22%.&nbsp; And I tried to get ChatGPT to write this script,&nbsp;&nbsp; but it still can’t tell relevant from&nbsp; irrelevant information. It has, however,&nbsp;&nbsp; mastered the important skill of sounding&nbsp; very confident while wasting my time.
The most telling failure though is&nbsp; Leopold Aschenbrenner’s prediction&nbsp;&nbsp; of the impending intelligence explosion by&nbsp; 2027. This, if you remember was supposed to&nbsp;&nbsp; happen by building a few 100 GigaWatt power&nbsp; plants by drilling for a little more gas.&nbsp;&nbsp;
And that in return was supposed to be done&nbsp; by robots who were built by the AI run on&nbsp;&nbsp; the gas-powered supercomputers that were&nbsp; built by the robots. You see the problem. That the Silicon Valley bros&nbsp; let this guy handle $30 billion&nbsp;&nbsp;
tells you how far gone they are. Because really&nbsp; the picture here isn’t so difficult to see. How&nbsp;&nbsp; fast you can change reality depends dramatically&nbsp; on how many physical changes you need to perform,&nbsp;&nbsp;
and how many institutional hurdles stand in the&nbsp; way. Maths, coding, and scientific publications&nbsp;&nbsp; are a straight entry with near zero resistance.&nbsp; But building a lab, building autonomous vehicles,&nbsp;&nbsp;
building robots, changing workflows in&nbsp; existing companies, building power plants,&nbsp;&nbsp; mining resources, etc. That’s a completely&nbsp; different thing. It will take decades. And from that we can also extrapolate&nbsp; what’s going to happen in the near future.&nbsp;&nbsp;
Rapid progress in anything that&nbsp; can be done by software alone,&nbsp;&nbsp; especially if that’s tasks which previously&nbsp; were impossible, such as reconciling all&nbsp;&nbsp; mathematics literature, which has now become&nbsp; possible. There are many theoretical problems&nbsp;&nbsp;
in physics and related disciplines that will be&nbsp; up for grabs soon. Because they are solvable.&nbsp;&nbsp; And the techniques for solving them are&nbsp; already in the literature. Maybe physics&nbsp;&nbsp; isn’t waiting for a new Einstein, it’s just&nbsp; waiting for someone to read all the papers!
The most exciting thing is that this is not&nbsp; exclusively something that AI companies can&nbsp;&nbsp; do. Everyone can contribute. And that, I think, is&nbsp; the future we’re headed into. The glorious phase&nbsp;&nbsp;
in which everyone is a scientist, for a few years.&nbsp; And then… Well, then we will either all die,&nbsp;&nbsp; or never work again, depending on whom&nbsp; you ask. I guess the lesson here is:&nbsp;&nbsp; don’t listen to the experts. Follow&nbsp; me for more uninformed opinions..
Artificial intelligence is developing rapidly, and&nbsp; OpenAI's GPT-6 Astra is closer to AGI than ever.&nbsp;&nbsp; Are you keeping up? Are you sure? Today's sponsor,&nbsp; Outskill, will bring you up to speed. Outskill's&nbsp;&nbsp;
offering a GPT-6 Astra crash course. It's 3 hours&nbsp; long, and it'll teach you how to build your own&nbsp;&nbsp; Astra workflow hands-on. Outskill is an official&nbsp; OpenAI partner, so their team has exclusive access&nbsp;&nbsp;
before anyone else. They've tested it upside down&nbsp; and had time to figure out how to get the most&nbsp;&nbsp; out of it. In the workshop, you'll learn,&nbsp; for example, how you can use Astra to find&nbsp;&nbsp; jobs that are tailored to your skills, or how to&nbsp; alert you of information concerning your stocks.&nbsp;&nbsp;
When you sign up, you'll also get the Astra power&nbsp; playbook with prompt formulas and workflows ready&nbsp;&nbsp; for you to plug and play. The link to sign-up's&nbsp; in the description, or you can scan the QR code on&nbsp;&nbsp;
screen. I've signed up for it because it's an easy&nbsp; way to stay up-to-date. So, go ahead, sign up for&nbsp;&nbsp; this, and learn what less than 1% on this planet&nbsp; know. Thanks for watching. See you tomorrow.
