Summary
Physicist and science commentator Sabine Hossenfelder looks at a recurring pattern in AI: experts keep underestimating how fast the technology advances. The video lines up forecasts from surveys of researchers against what actually happened, from solving a Millennium Prize problem (median expert estimate of 10% by the end of 2027) to the LiveCodeBench Pro programming benchmark, where the best score of 7.7% was expected to reach only 14% by the end of 2026 but hit 53.8% by May.
She also covers the AI task completion time horizon, which measures how long a coding job a model can reliably finish. Experts predicted about 3.4 hours by the end of 2026, and a new model had already reached 3 hours and 6 minutes while the survey was still running. Earlier misses include autonomous vulnerability discovery in cybersecurity, International Math Olympiad gold-level performance arriving in 2025 rather than 2030, and a flood of AI-written papers on physics and maths preprint servers.
The second half examines where predictions went the other way. Forecasts of rapid job automation, including the World Economic Forum’s expectations and the early replacement of truck drivers, have largely not materialized, and Leopold Aschenbrenner’s predicted 2027 intelligence explosion is criticized as unrealistic. Hossenfelder argues that software-only fields like maths, coding and scientific publishing face little friction, while anything requiring labs, robots, power plants or changes to institutions will take decades.
📺 Source: Sabine Hossenfelder · Published October 07, 2026
🏷️ Format: Opinion Editorial
