> So my question is, how hard of a problem is that with LLMs? I get the sense that LLMs are trained on a very large data set all at once, but that it is difficu

> So my question is, how hard of a problem is that with LLMs? I get the sense that LLMs are trained on a very large data set all at once, but that it is difficult to incrementally update them with new data. Is that a true assessment? That's not quite correct. It's trivial to update the model with new information. What's not trivial is to encode priorities or time into the model. Fundamentally, the model doesn't understand about the concept of time, so you can't easily condition it on "give me only information from last year". You can get there and nudge it in the right direction by generating training datasets for such queries and fine-tuning the model on them, by oversampling more recent data in your training procedure, or by adjusting your gradients for more recent data, but all of these are still "fuzzy" - the model may or may not do exactly what you want during prediction time and could still give you old results. However, I would assume that Google is doing something more sophisticated and that's it's not just a plain LLM like GPT3. It's probably a more complex architecture with external systems and data sources around it that tries to solve the above using heuristics. That being said, the current Google search is extremely bad at time relevance too...