That can be true if it is using “tools” [1] and/or retrieval augmented generation. Something doesn’t have to be in the training set for it to be returned to you and used in generation as long as the model knows that a particular tool will be useful in responding to a particular prompt.
[1] This is what people call plugins that provide additional context to a gpt model
They (Google) are probably using tools in a different way. I would imagine if you ask Bard/Gemini something, it also does a google search at the same time and provides those results as a potential context that the chat bot can use to answer with. So it does a google search every question but doesn't always use it.
With chatGPT it only uses the tools if it thinks it needs it. So if it needs to do a search it will have to respond with do a search function, which then has to go do a search and then it provides that as context to the chatbot which then can respond from that data.
I think this is possibly true, but if it is, it blows GPT-4s use of "tools" out of the water. GPT4 browsing the web is much slower and doesn't feel as well-integrated. It feels about the same speed as me opening the page myself and reading it. Whatever Gemini did, it was significantly faster.
I don't know how they've specifically done it, either, but this is an area where Google has a ridiculous advantage over pure play AI shops. It's highly likely they have architected it for use cases like this from the outset, since the primary application of Gemini will be within Google's own products. They'll publish APIs, of course, and embed within Vertex AI on Google Cloud, but since the primary utility of Gemini will be to improve Search, Maps, Travel, Youtube, etc, I'd imagine they had a first class business requirement from the beginning along the lines of "must be easy to plug into existing Google data sources & products."
When Bard inserts that information unasked (as in something like "I'm sorry but I don't have that information due my training data cutoff being ...") It may quote other later dates. I got a response with "October 2023" at least once so far.
