Weekly Dwell #3
In Snow Crash (yes, the one about the metaverse), Hiro Protagonist goggles into the metaverse and accesses a librarian who then helps him piece together everything that is happening with the snow crash drug. This librarian is a form of AI, who helps bring up information as he is searching, drawing connections with him but also understanding their limitations about drawing conclusions or causation. This week, Meta released their large language model (LLM) Galactica as a way to help scientists (social and mathematical alike) do research better, with the end goal of potentially replacing search engines entirely by having instead conversations with an AI that could do a similar type of search for the correct information. However, as with most Web3 related endeavors, it didn't go well - providing instead biased and unscientific information.
According to an article from MIT Technology Review, "Galactica is a large language model for science, trained on 48 million examples of scientific articles, websites, textbooks, lecture notes, and encyclopedias. Meta promoted its model as a shortcut for researchers and students. In the company’s words, Galactica “can summarize academic papers, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more.”" This sounds like the right data set, with the right potential outcome. What could go wrong?
Mainly, people were asking it questions that it tried to find the answers to the best it could, linking causation when that is not really an ability it has. Unlike the Librarian in Snow Crash, who would send a message saying that they could not tell the "why" of something, but rather could bring all the information to light in hopes of the human reader to make sense of it, Galactica aimed to answer "why" questions that are just out of the scope of its ability (as it very well should). Using an AI as a librarian can be helpful in some ways (this is, essentially, what Google currently does, using an algorithm to match best results to the query), but certainly has limitations and should not be used lightly or really for anything beyond pulling up full information.
Personally, the idea of an AI librarian, which might be able to assist actual librarians rather than laypeople, could be a helpful tool in the instance of drawing relevant data together. Actually using an AI to build any form of text (beyond pure satire) that makes conclusions or contains analysis is, however, something far more dangerous and discouraging. It will not only work to eliminate human librarians, but also human researchers, and therefore start to question the extent in which knowledge and the pursuit of it is human at all.