[scandinavian]

> OK, so if google is training bard in real time

Is that what is happening here? I assumed that both bing and bard worked something like this:

* Take the user question and run it through a GPT like model to generate a search query.

* Search google/bing with the search query and get the top-x results.

* Take the results and append the original user question with some additional prompt glue like "Given the following search results: SEARCH_RESULTS - Generate an answer to the following question: ORIGINAL_QUESTION.

I'm not sure it's technically feasible to live-train any of these models.