DIY interpretability: polynomial fitting plus coin-flip RLHF to explore how features explain language

cephaloform · x · 2026-09-16

A developer shares a toy experiment in language explainability: instead of training a big model, fit a polynomial to a hand-scored set of words, use a Markov chain to decide part of speech, and find words closest in the feature dimension.

He also shows how to do easy "RLHF": for each weight, flip a coin for +0.1 or -0.1, generate from the updated model, keep the update if better, negate if worse — random exploration as a stand-in for preference optimization.

Related event: Dev demos pen-and-paper semantic modeling with polynomial fits and Markov chains(5 posts)→

Original post →

More from Fun

Fun channel →