Strong test-time learning seems to be both an AGI crux and makes open weight models highly capable in arbitrary domains. Narrow capabilities developed post-deployment, at costs comparable to inference when working on corresponding narrow tasks, make the pre-deployment level of capabilities less important. For AGI-level adaptation, this should extend to capabilities such as playing chess, designing biological gadgets, or becoming fluent in a novel area of math, all at least at the level of the most proficient humans, even if the pre-deployment version of the model couldn't do those things at all.

If this is an algorithmic innovation, the destabilizing AGI threshold that makes true RSI feasible is crossed precisely when open weight models can also cross it, regardless of which capabilities are available in them pre-deployment, or what guardrails govern the closed models. This doesn't happen only if strong post-deployment adaptation requires some very compute-intensive pre-deployment step similar to pretraining or RLVR on a broad range of tasks, and it wasn't applied to any open weight models available at the time, before the world wakes up to the implications of strong post-deployment adaptation.