Has there been research on neural nets who could cite their sources? In common cases, thousands or millions of items in the training set have contributed to the weights in the pathways that activate for a certain input. But maybe for some inputs, only a small number of items in the training set contributed strongly to the weights on the pathways the are decisive to the output? Could you design a data structure that would keep this information in the neural net?
I wonder if this is a special case for Explainable Models [1] — once you select an answer, another model helps describe and source the output.
[1] https://en.m.wikipedia.org/wiki/Explainable_artificial_intel...
I believe it's going to be in Bing soon?
I am going to guess that even though it uses the GPT language model, it works fundamentally differently than ChatGPT.
My guess is, Perplexity AI is a traditional search engine, that first finds a list of links that match your questions. And then it feeds text from those links to the GPT language model to write a summary of the results.
Whereas ChatGPT produces text based only on the trained deep neural net.
Maybe I am guessing wrong how Perplexity AI works. Maybe it works in a different manner.