They learn again when the labs train the next version. The previous version often has an indispensable role in orchestrating that process, as a critic and from

They learn again when the labs train the next version. The previous version often has an indispensable role in orchestrating that process, as a critic and from usage transcripts of the previous model that they have access to. And between two training runs the model can assemble a lot of md files, memory files, and load them in on-demand. I think this is enough to make it scifi-level, but I agree there should be some way to customize it at a deeper level and adapt to each person's context. But things are still moving too fast for that to make sense. If they fine tuned a specific LoRA for Claude just for your projects and workflows, it would get outdated quite fast and you'd have to retrain it at quite a cost for the new version.