Hutter et al. Argue Generalization Requires Universal Induction in New arXiv Paper
examachine · x · 2026-10-03
A new arXiv paper by Aram Ebtekar, Marcus Hutter, and Danica J. Sutherland argues classical statistical theory can't explain general-purpose AI: No Free Lunch theorems imply any learner beating chance somewhere must lose elsewhere, and this applies to meta-learning too, so meaningful prediction always starts with an inductive bias external to the data.
Biasing toward short programs yields Solomonoff induction (SI), which the authors relativize to an information vantage point—favoring accessibility relative to preexisting information rather than absolute simplicity. Any algorithm outpredicting relativized SI must embed extra information about the data, which none can generate. SI is incomputable, but it formalizes the inference optimum at infinite compute, with evidence suggesting frontier AI systems roughly approximate it.
More from AGI Musings
- Yann LeCun at ETH Zürich: Scaling LLMs to reach AGI is 'impossible' — rohanpaul_ai · 2026-10-03
- VC accuses OpenAI of skynet-style regulatory capture, says Anthropic's version is more subtle — StewartalsopIII · 2026-10-03
- Anil Seth pushes back on Steve Yegge: denying machine consciousness isn't human exceptionalism — anilkseth · 2026-10-03
- Redditor predicts AI bubble pops within 1-2 years: bond markets aren't charities — lewispatty · 2026-10-03
- DeepMind researchers propose Artificial Symbiotic Intelligence over singularity — The Decoder · 2026-10-03
- Avi Loeb: One non-human UAP in a million would matter more than AI — examachine · 2026-10-03