The approach of LLMs without some sort of symbolic reasoning layer aren't actually able to hold a model of what their context is and their relationships. They predict the next token, but fall apart when you change the numbers in a problem or add some negation to the prompt.
Awesome for protein research, summarization, speech recognition, speech generation, deep fakes, spam creation, RAG document summary, brainstorming, content classification, etc. I don't even think we've found all the patterns they'd be great at predicting.
There are tons of great uses, but just throwing more data, memory, compute, and power at transformers is likely to hit a wall without new models. All the AGI hype is a bit overblown. That's not from me that's Noam Chomsky https://youtu.be/axuGfh4UR9Q?t=9271.
https://quantian.substack.com/p/market-prices-are-not-probabilities#footnote-anchor-1-150368734