Replying to
@Urban_Hermit @spaf I absolutely agree, though some notes on those points:
Checking our thoughts for consistency / logic: we've got a basic form of that with "reasoning" models, they do a round of thinking to sort things out before they actually reply.
And as far as forgetting everything, we've got Retrieval Augmented Generation, which basically gives them long term memory outside of training data. It's not part of the model itself, so it's not part of the APIs, but it's built into interfaces like ChatGPT and is why it can remember things about you between conversations or even just hold a long continuous conversation without hitting context window limits.
Basically it boils down to giving them a system of short term (the context window) and longer term (the RAG system) memory where the RAG will cycle things in and out of the context window as they come up.
Though I don't know how much this would qualify as actual learning vs just learning facts. My money is on more facts than actual learning, but it's harder to tell when it's restricted to just language.
I believe Tav was made with actual dynamic training, which was a horrible idea at scale.
I do think we'll get there with incremental advancement rather than revolutionary, but I do agree it will be a very different beast from current LLMs. Especially given that they would probably need more complex motivations.