I believe the Code4Lib conference was an unqualified success, kudos to the organizers at the University of Michigan, and the agenda included more than a few presentations/workshops on the topic of AI.
Along similar lines, I have done some tertiary analysis against the queries/responses submitted to my chatbot experiments. On a temporary basis, you can see the transcripts here:
https://distantreader.org/tmp/chat-logs/
Based on my initial observations, folks came to the experiments with expectations, both realistic and unrealistic. On the realistic side, a number of people expected the chatbots to maintain context. In other words, people tried to carry on a conversations with the systems believing the inputs & outputs would be included in subsequent queries. Unfortunately, that's not how these 'bots were configured.
On the unrealistic side, some people seemed to expect the underling sytem to have indexed a very wide -- if not complete -- set of content on a given subject. This was not the case. In every case, there were fewer than a couple hundred articles / book chapters vectorized underneath. Moreover, these indexed items may have been older rather newer, and consequently they would not be able to address things that happened in the recent past nor outside the scope of the content.
We are suffering from information overload. Such has been the case of at least a millenium. Many techniques have been posited as solutions, and I believe the implementation of RAG (retrieval-augmented generation) 'bots presented as subsequent parts of search interfaces may be the newest. Do a search. Select items of interest. Have just those items indexed. Query the result. Repeat. Bing's Co-Pilot is an example.
--
Eric Morgan <[log in to unmask]>
University of Notre Dame
|
