Schmidhuber: Physical AI progress is overestimated, human-level physical intelligence not within 2 years

SakanaAILabs · x · 2026-11-10

Sakana AI Chief Scientific Advisor Jürgen Schmidhuber gave an interview to Nikkei and NIKKEI Digital Governance on the path beyond LLMs.

Key points:

"Even if you become world champion at chess, you can't do a plumber's job" — today's LLMs excel at solving questions learned from web-scale data, but human babies build world models by running their own experiments and generating data stored on no server
Working AI today only lives "behind the screen"; Physical AI needs new kinds of data for understanding the real world
He believes Physical AI progress is overestimated: human-level intelligence in the physical domain won't arrive "within 2 years"; the key question is building general robots that learn to operate any machine or tool without being overwhelmingly smart
Japan (home of Kunihiko Fukushima) and Germany could regain prominence through Physical AI
On safety, his main worry is human conflicts of interest being imported into AI, making globally meaningful rules hard to establish

More from AGI Musings