Schmidhuber rebuts 2028 science-AGI claim: physical AI will take far longer
On September 23, AI pioneer Jürgen Schmidhuber posted a series of replies in the debate over whether AI can surpass all sciences by 2028, pushing back on optimistic timelines associated with Elon Musk and François Chollet. His core argument: every truly effective AI today lives in the virtual world behind screens, while physical-world AI difficulty is badly underestimated.
Confirmed
- He reiterated his long-held view that today's useful AI handles virtual tasks—document summarization, chess, theorem proving, image, code and slide generation—while no robot has shown real mastery of the physical world, and without such mastery there is no AGI.
- He argued the Turing test is not a measure of intelligence, and that real science largely involves learning to interact with the physical world, where current fragile AI-driven robots perform poorly.
- Using self-driving as an example, he noted autonomous cars are 'very simple robots' with only 2 degrees of freedom—steering and acceleration—despite massive inputs, citing Dickmanns' vision-guided car exceeding 170 km/h on a highway back in 1994.
- He pointed out that compute became roughly a million times cheaper over 30+ years, yet robotics and self-driving capabilities did not improve a millionfold—evidence that Physical AI is slow work: even this simple problem remains troubled 32 years on.
- As he has predicted for fifty years, superhuman physical AI will eventually arrive, but 'much more than two years' is needed, and near-term timelines are exaggerated.
Why it matters
- The dispute cuts to the heart of current AI-route disagreements: generative AI in virtual domains advances rapidly, while embodied intelligence and physical interaction remain weak. Schmidhuber's counterexample—that even self-driving with minimal degrees of freedom remains unsolved—pours cold water on radical timelines and bolsters his long-standing physical-AI research agenda.
2026-09-23 ~ 2026-09-23 · 6 related posts
Primary sources
- Schmidhuber: Turing Test is a bad intelligence measure — no AGI without mastery of the real world — SchmidhuberAI · 2026-09-23
- Schmidhuber: superhuman physical AI will come, but not within 2 years — SchmidhuberAI · 2026-09-23
- Schmidhuber: self-driving cars since 1994 still unsolved — physical AI is a slow game — SchmidhuberAI · 2026-09-23
- [source] Schmidhuber rebuts Musk's AI timeline: no robot can match a plumber, let alone beat science by 2028 — AjdDavison · 2026-09-23
- Schmidhuber: compute got 1M times cheaper, robots and self-driving did not — SchmidhuberAI · 2026-09-23
- Schmidhuber: no, AI won't beat all fields of science by 2028 — physical world is far harder — SchmidhuberAI · 2026-09-23