In thinking of directions where LLM's could develop from here, I cant help but think that a models ability to self introspect would immensely improve their utility. The R&D on how to achieve that is beyond me though. How do you train someone how to introspect? Also would it require a continuous learning architecture that doesn't separate training and inference?

I think it’s possible this might just require reworking the transformer architecture a bit to give it an inherent awareness of the contents of its attention. Kind of a working memory workspace of sorts.

I.e. Instead of limiting autoregression to the token level, you introduce a persistent compressed global workspace latent memory vector that is fed back into the self-attention mechanism at every layer or every token step, allowing the network to attend to its own prior attentional states before computing the next token. Obviously that’s going to involve some compression steps.

Trouble is… I think the architecture there is much simpler a tweak than figuring out how to train it.

…that’s likely to just destabilize training for not much if any gain at first. You’re probably gonna have to resort to some really clever (and currently missing) tricks to figure out how to train the network to actually use that feature.

Well, the algorithms in social media that are rewarding "engagement" (but accidentally ripping society apart by fueling rage) have made the training data — online discussions — a very poor training medium for introspection.

So, if companies want training data on introspection, they would need to first fix what they broke (trust in each other and society) and have those discussions and realizations made public (by bringing back blogs where people might jot down their "maybe I've been an asshole" realizations) and rewarding other things than "engagement".