Disclaimer: I've done a lot of stuff with local models and RAG methods, I haven't done a lot of work with public models and so don't know how Gemini, GPT and so

Disclaimer: I've done a lot of stuff with local models and RAG methods, I haven't done a lot of work with public models and so don't know how Gemini, GPT and so on are working right now. Claude + GraphRAG through Bedrock is my main mode of playing with this stuff right now. Things LLMs are good at include summarisation and contextualisation. They can use that facility to help summarise processes and steps to get something done, if they've been trained on lots of descriptions of how to do that thing. What they're not good at is perfect recall without being nudged. This example would have been very different if the LLM had been able to RAG (or GraphRAG), a local data source on palaeontology. I think we're going to see an evolution where search companies can hook up an LLM to a [Graph]RAG optimised search index, and you'll see an improved response to general knowledge questions like this. In fact, I'd be surprised if this isn't happening already. LLMs on their own are a lossy compression of training material that allow a stochastic parrot to, well, parrot things stochastically. RAG methods allow more deterministic retrieval, which when combined with language contextualisation can lead to the kinds of results the author is looking for, IME.