This is called Test Time Learning and some research architectures can do that. Current Mainstream models may not do that because their design is mostly about sc

This is called Test Time Learning and some research architectures can do that. Current Mainstream models may not do that because their design is mostly about scalability. They have to serve millions of people with low latency. Alternatively they could design and run a single super-intelligent model, with no scalability constraints. Probably whey are already doing that as well.