> This learning happened without changing the underlying model weights, but through experience accumulated in editable text files.

I think this is why I wouldn't call it play, but it's certainly pretty convincing anyway. We just need more research into AIs that change their weights at inference time, or perhaps a different architecture that requires the AI to be "always on".