Oxford researchers argue LLMs can't invent: 'Theory Is All You Need' paper on data vs. human cognition
alex_verem · x · 2026-11-08
Teppo Felin (Utah State) and Matthias Holweg (Oxford) argue in Strategy Science that LLMs, being probability-based predictors over past data, gravitate toward majority views and can rarely invent. A model trained on all texts up to 1633 would call Galileo delusional. Flight history echoes this: Kelvin dismissed aerial navigation in 1896, the NYT put flying machines 1–10 million years away in October 1903, and the Wright brothers flew 69 days later by decomposing flight and building their own wind tunnels. Training data (13T tokens) would take 164,000 years to read, yet a child hears 36.5M words and still says novel sentences. The authors suggest AI for routine decisions, caution on decisions with no data yet, and warn that companies querying the same model will converge on identical answers.
More from AGI Musings
- 3,000 years of frozen family careers: why abundance, not money, expands life choices — yunta_tsai · 2026-11-08
- Box CEO Aaron Levie: AI Has Widened the Engineer Gap to at Least 100x — 2C_ornot2C · 2026-11-08
- Academic protest: reviewers caught using AI to auto-generate peer reviews — TuhinChakr · 2026-11-08
- Dev calls for AI's 'minister of propaganda' to fight anti-AI sentiment — zack_overflow · 2026-11-08
- Anders Sandberg: the real challenge is understanding where and why progress levels off — anderssandberg · 2026-11-08
- ChrisGPT: We could automate most work in 10 years, yet fear-driven bans dominate — ChrisGPT · 2026-11-08