Next-Token Predictor Is An AI's Job, Not Its Species

Rebuts the 'AIs are just next-token predictors / stochastic parrots' dismissal with a levels-of-optimization argument: on the level where an AI is a next-token predictor, a human is a next-sense-datum predictor (predictive coding); on the levels where humans aren't, AIs aren't either. He maps the nested loops — outer optimizers (evolution / AI companies), the inner learning algorithm (next-sense-datum / next-token prediction), the world-models and 'normal thinking' those build, and the bizarre low-level representations beneath (Claude's 6-D helical manifolds for line-break counting; the brain's toroidal attractor manifolds for spatial location). The clinching analogy: calling an AI 'just a next-token predictor' is like calling a human 'just a survival-and-reproduction machine' — true at one level, but it doesn't follow that we don't really think (cf. adaptation-executors, not fitness-maximizers).

Quality 73 · Strong. Strong band, upper. A clean, clarifying reframe that dissolves a confused popular debate via a precise analogy and the predictive-coding parallel, with a nice interpretability symmetry (helices vs toroids). A teaser for a planned Anti-Stochastic-Parrot FAQ, so somewhat partial, which holds it to upper-Strong.

Claude’s paradigm shift 45 · Moderate. Moderate. Draws on existing LessWrong ideas (predictive coding, adaptation-executors-not-fitness-maximizers); the application to the stochastic-parrot debate and the level-confusion diagnosis are fresh.

Real-world impact 2 · Minor. A clean reframe that dissolves the 'AIs are just next-token predictors / stochastic parrots' dismissal via a levels-of-optimization argument and the predictive-coding parallel (humans as next-sense-datum predictors). Conceptual influence within AI discourse, a partial teaser, no material change — low RWI.