As of: 2026-10-09 19:24 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code). Human editor: Adam Bicz.
Canonical page: https://postcutoff.com/v/sabine-experts-were-wrong-about-ai-again/

# The Experts Were Wrong About AI. Again.

Sabine Hossenfelder, 7 October 2026, YouTube. 582,176 views as of 9 October 2026.
Kind: Review.
Watch: https://www.youtube.com/watch?v=xdwFncdu8t0

## Why it is here

Sabine Hossenfelder on how expert AI forecasts keep being wrong. ~581k views by 2026-10-09. Length 7:22.

## Description (written by Gemini from the video)

**Summary**  
Theoretical physicist and science communicator Sabine Hossenfelder analyzes how expert forecasts have systematically underestimated software-centric AI milestones while overestimating physical automation. She contrasts rapid advances in automated mathematics, coding, and cybersecurity against the physical and institutional bottlenecks facing robotics, autonomous driving, and massive power infrastructure, before presenting a sponsored overview of an Outskill workshop for GPT-6 Astra workflows.

**What is shown**  
- [00:11] A *Washington Post* headline from September 23, 2026: *"Using AI, a professor wrote 200 papers this year. Researchers are alarmed."* regarding University of Chicago professor Nicholas Polson.  
- [00:40] Forecasting Research Institute's *"Wave 2: AI for Science"* survey report (released November 10, 2025) showing predictions on the Millennium Prize Problems.  
- [01:10] Terminal screen showing coding lines and LiveCodeBench Pro prediction stats (status 7.7%, prediction 14%, actual result 53.8%).  
- [01:38] METR Time Horizon charts tracking benchmark performance and task completion horizons across models including GPT-4o, Claude 3.5 Sonnet, o1-preview, and Claude Mythos Preview.  
- [02:03] The January 2024 preprint *"Thousands of AI Authors on the Future of AI"* by Katja Grace et al. and its milestone timeline chart highlighting math theorem predictions.  
- [02:21] Post on X by Thomas G. Dietterich from September 13, 2026, describing arXiv flooded with AI-written papers authors don't understand.  
- [02:43] CSET 2023 report *"Autonomous Cyber Defense: A Roadmap from Lab to Ops"* alongside the arXiv paper *"LLM Agents Can Autonomously Hack Websites"*.  
- [03:26] Milestone graph highlighting truck driver automation projections, followed by footage of a truck driver.  
- [03:44] World Economic Forum *Future of Jobs Report 2025* chart on the shifting human-machine frontier (automation vs. augmentation).  
- [04:01] Leopold Aschenbrenner's *"Situational Awareness"* intelligence explosion scenario chart, followed by a cycle diagram of data centers, power plants, and robots.  
- [04:38] Aerial drone footage of chip fab/data center construction sites and cooling infrastructure.  
- [06:21] Promotional visuals for Outskill’s 3-hour live workshop on building OpenAI GPT-6 Astra workflows, showing job tailoring and financial monitoring interfaces, along with an on-screen signup QR code.

**Claims & numbers**  
- The presenter notes that on September 23, 2026, an online platform removed 257 papers written or co-written by Nicholas Polson after scrutiny over AI generation.  
- In a late 2025 Forecasting Research Institute survey of 277 experts, 58 superforecasters, and 1,022 members of the public, the median forecast estimated a 10% chance AI would solve or substantially assist in solving a Millennium Prize Problem by end-2027; the presenter claims AI reached this benchmark in September 2026.  
- On LiveCodeBench Pro, where the baseline was 7.7%, experts predicted a rise to 14% by end-2026; actual performance reached 53.8% by May 2026.  
- On METR task completion time horizons, experts surveyed in April/May 2026 predicted models would reach ~3.4 hours by end-2026; during the survey, a new model reached 3 hours and 6 minutes (with Claude Mythos Preview early reaching 17 hours).  
- In late 2023, Katja Grace et al. surveyed over 2,000 AI researchers who estimated it would take 22 years for AI to prove publishable mathematical theorems in top journals.  
- In 2022, experts predicted gold-level International Mathematical Olympiad (IMO) performance would arrive by 2030, but AI achieved it in 2025.  
- In 2020, the World Economic Forum predicted nearly half of work tasks would be automated by 2025; a 2025 follow-up found the actual rate was ~22%.  
- Outskill is offering a 3-hour live workshop at 10 AM EST on building workflows with OpenAI's GPT-6 Astra, including access to prompt formulas and the "Astra Power Playbook".

**Notable quotes**  
- [00:18] "We've gotten so used to AI so quickly that it's easy to forget just how much it's outperformed expectations."  
- [04:39] "How fast you can change reality depends dramatically on how many physical changes you need to perform, and how many institutional hurdles stand in the way."  
- [06:01] "I guess the lesson here is: don't listen to the experts."

**Assessment**  
This is a commentary and science news analysis piece presented directly to camera with overlaid paper screenshots, charts, and b-roll, concluding with a sponsored pitch for a training workshop. The presenter cites verified academic surveys, benchmarks, and reporting to support her argument, but does not conduct live AI testing herself.

_Described by gemini-3.8-flash on 2026-10-09 from the video's audio and frames._

## People in it

- [Leopold Aschenbrenner](https://postcutoff.com/person/leopold-aschenbrenner/), Founder, Situational Awareness LP