Comment by lynndotpy

Others have explained the confusion, but I'd like to add some technical details:

LLMs are what we used to call txt2txt models. The output strings which are interpreted by the code running the model to take actions like re-prompting the model with more text, or in this case, searching Twitter (to provide text to prompt the model with). We call this "RAG" or "retrieval augmented generation", and if you were around for old-timey symbolic AI, it's kind of like a really hacky mesh of neural 'AI' and symbolic AI.

The important thing is that user-provided prompt is usually prepended and/or appended with extra prompts. In this case, it seems it has extra instructions to search for Musk's opinion.