Anthropic trains models that take neural activations as input and report on them — surfacing knowledge and misalignment that fine-tuning had hidden. A model examining its own weights is a different kind of recursion from one watching its training curves, and the corpus now holds both.
The rest of the issue is generalism winning: models learning from Twitch footage, late-blooming generalists overtaking prodigies, specialization losing its premium everywhere at once.