As long as they are static they wont be conscious. And once they are dynamic we wont call them transformer architecture anymore, as the dynamic part is the important part at that point.
Maybe we'll call it "continuous RLHF" or something like that.
But you might be right that the dynamic part might be the biggest architectural shift needed. You can simulate a lot with in-context memory or clever retrieval, but memory alone doesn't allow the model to get better at chess the same way a human does
But you might be right that the dynamic part might be the biggest architectural shift needed. You can simulate a lot with in-context memory or clever retrieval, but memory alone doesn't allow the model to get better at chess the same way a human does