[koakuma-chan]

Not really. For example we still can’t get coding agents to work reliably, and I think it’s a memory problem, not a capabilities problem.

[atlex2]

On the other hand, test-time weight updates would make model interpretability much harder.