3 ms·

Anyone here up on AI training in this sense? Example, google certainly indexes search fast. I have googled a product posted 10 mins ago in a very low-key spec

Anyone here up on AI training in this sense? Example, google certainly indexes search fast. I have googled a product posted 10 mins ago in a very low-key specialized audio forum that is really not popular, and bam, google is liking to the post. OK, so if google is training bard in real time, could it not encounter the classic text book example of training on future data? e.g. we talk about it, or what it might or might not do in HN, Twit, (God help us some meta-universe), or elsewhere, and it is training on that as factual input? What then? Tail wagging the dog? What mechanisms are in place to make that be noise and not facts, I wonder.

scandinavian 4y ago
> 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.