'Just Next-Word Prediction' Massively Underrates AI: Predicting Proof-Level Tokens No Human Can Match
jeremiecharris · x · 2026-11-19
Jeremie Harris endorses an argument that 'it just predicts the next word' massively undervalues how hard that task really is.
The quoted argument: predicting the next word in a mostly finished sentence is easy, but predicting every next word in a math proof well enough to solve Navier-Stokes — as GPT 6.2 Astra just did — is something no human could do. The MMA analogy: a fighter who predicts your next strike would go undefeated in the UFC even with below-average strength and speed.
Harris adds that while academic, it's a good intuition pump for why AI can go further than many think — even filling in 'To prevent China from stealing America's best models, the U.S. should ' is 'just autocomplete'.
More from AGI Musings
- wordgrammer essay: passion can't be outcompeted — great talent unifiers beat peer-group sorting — wordgrammer · 2026-11-19
- Yoav Goldberg: literary experts dislike AI prose less than we assume — yoavgo · 2026-11-19
- Yoav Goldberg on AI-writing detection row: disliking AI style isn't an accusation — yoavgo · 2026-11-19
- LOTR meme mocks the 'grab the ring first' AI race logic — peterwildeford · 2026-11-19
- AI makes 'living papers' nearly free to build, but their durability is doubtful — _akpiper · 2026-11-19
- Ghent professor: AI-era PhDs produce results they can't explain; Goldman argues journals are obsolete — 2C_ornot2C · 2026-11-19