This adds to my feeling that LLMs are incredible at storing compressed information and accessing it in a versatile way. Apart from occasional hallucinations we

This adds to my feeling that LLMs are incredible at storing compressed information and accessing it in a versatile way. Apart from occasional hallucinations we celebrate their ability to dissect information and to find obscure links between facts. Hence their stunning effectiveness in finding security holes or new approaches to math problems. Interestingly the generative part - producing the answers - is what we are criticizing a lot: we think the code is verbose, texts have a weird repetitiveness to them etc.