> On a serious note, it seems to me technology isn't that far from that point.

It feels at least a couple of decades away.

I think the most compelling use for LLMs and generative AI right now is parsing massive tomes of knowledge, which can then be summarised at use-time.

Like imagine pilots being able to consult an LLM with all of the aircraft's blueprints, technical reports, operational history, etc. during flight. Likewise for spacecraft when debugging in flight.

But the internals really need to go beyond just token prediction, to being able to use Abstract Syntax Trees for programming suggestions with correct types, or building ontology trees for answering queries factually. The former is a major frustration with Copilot (when it gives relevant suggestions at all).