[wongarsu]

We would presumably stop calling it an LLM somewhere along the way. But I don't see why it couldn't be a transformer architecture at the heart of it, and why that transformer couldn't bee pretrained from Reddit. You would have to track a lot of stuff on to allow an internal stream of consciousness and interaction with the world as well as memory, and do significant reinforcement learning. But we are already doing all of that while still calling the thing an LLM. It's unclear to me where the border lies where it would cease to be an LLM

[Jensson]

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.

[wongarsu]

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