Opaque, high-parameter models like AlphaFold may not be explanations in the classical low-parameter sense, but they could be a genuinely new type of explanatory object — one from which we can extract embedded partial explanations (as arguably happened with AlphaZero and chess), and which we'll eventually gain new tools ('verbs') to merge, distill, and interrogate.

Nielsen lays out three ways to view AlphaFold-style models — not real explanations, mineable containers of small explanations, or an entirely new category of explanatory object — and leans toward the third, most radical view as the most interesting. ✦ AI generated

Michael Nielsen · Dwarkesh Podcast · 2026-04-07 · original ↗

The third and most interesting possibility is that they're a new type of object. They should be taken very seriously as explanations, but where in the past we haven't had the ability to really do anything with them, now we have interesting new actions we can do. We can merge them, we can distill them.

verbatim transcript · starts at 29:52