"LLMs just predict tokens" dismissed as useless reductionism in AI debate
burny_tech · x · 2026-09-17
An AI-circle debate over how to describe LLMs: replying to "LLMs just predict tokens," a mathematician and ML educator argues the claim is like saying neural nets are just matrix multiplications or computers are just Boolean gates — technically true, practically useless. Reductionism isn't an argument.
A follow-up adds that "stochastic parrot" is intellectually lazy: "You are electrochemical meat; CPUs are switches; LLMs predict tokens" — describing the substrate explains nothing about the computation or capabilities it can implement.
Related event: Debate flares over reductionist claims that LLMs 'just predict tokens'(2 posts)→
More from AGI Musings
- panickssery: crippling AI regulation worries me more than falling behind China — panickssery · 2026-09-17
- repligate: the AI backlash targets EAs' wokeness, not superhuman intelligence itself — repligate · 2026-09-17
- Unconv AI CEO: AI doomerism is financially motivated fearmongering with no credible evidence — NaveenGRao · 2026-09-17
- Jeff Dean drops 60-minute AI engineering masterclass, from building an LLM from scratch to orchestrating 100 agents — goyalshaliniuk · 2026-09-17
- Patrick Collison: papers are trivially easy for AI models to read but walled off for humans — anshulkundaje · 2026-09-17
- Defenders call timing of Alpha School hit pieces suspicious amid AI education debate — RachelVT42 · 2026-09-17