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.
> Overall, our results indicate that current language models possess some functional introspective awareness of their own internal states. We stress that in today’s models, this capacity is highly unreliable and context-dependent; however, it may continue to develop with further improvements to model capabilities.
3 comments