I’ve been reading a lot and playing around a little with ChatGTP and Bing, guided by some helpful guest speakers at work. On the whole, I’m a curmudgeon, though not for reasons related to my sci-fi fandom or comp sci degree. Instead, I am leery as a reader because I do think these large language models (LLM) value answers but not thoughtful or factually correct writing. More selfishly I’m also risk averse and setting aside larger fears I do think it will be disruptive to my field and others reliant on writing.
But there’s no running from automation. I think this effect will be mitigated by real constraints on time to read. Quality and trustworthy writing is more rewarding for readers at the same time cost, though I imagine this will strengthen winner take all dynamics and the importance of brands as the internet faces an ever growing tide of copycat mass-produced writing.
My biggest challenge as a scholar is the way LLM default to breaking the chain of custody for primary sources and ideas. This is a bit like wanting to identify the source tree for an individual bite of apple sauce, though Bing at least is experimenting in being able to cite. I think Tim Hickson does a good job of unpacking a range of issues in the AI art domain. One point that stood out for me is that Adobe’s generative art program and the Stable Diffusion music generator are explicitly limited to sources where they have rights due to public domain or arrangements with creators.
The scholarly dynamics for research and theory are different from art, although there are elements of both thinking and craft to unpack. I am glad not to be a teacher in this environment, because I find Paul Musgrave and Alan Jacob’s arguments persuasive and depressing. At the same time, I do think Dan Nexon is right to experiment with what the tools can and can’t do as well as thinking about how the demand for writers will change in each field.
So what are the opportunities and where does it add the most value? Starting with the big picture, there’s already more written material produced constantly than any of us will ever have a chance to read and moreover writing is a favored form of human expression that people make sacrifices to engage in. Same for visual arts and music. That said, customization is key; there are lots of places where people want middling writing or art: all sorts of promotional materials, a summary which zeroes in on a certain element of a large corpus of writing, a picture of your roleplaying game character. Quality is nice in these cases but a 33 percent solution at your finger tips could greatly increase demand.
In my work, I see three areas that are particularly promising:
- Low grade translation to identify promising documents. Credit to my colleague Alexander for this idea, but searching and summarizing documents with low fidelity can help us identify where to engage humans for high fidelity work. Timothy Lee recently wrote on how AI is changing the field of translation.
- Supervised categorization is not new to LLM, but may ease the utilization of large text fields in the process. This may enable myself and colleagues to put more work into defining taxonomies and adjudicating border cases and less work on easy but numerous calls.
- Search and summary within a particular corpus, in particular budget and oversight reports or policy documents or contract descriptions. Here AI may be able to do some scouting and I could then use that to inform my subsequent engagement with the primary sources.
All three cases involve existing applications of machine learning and other software tools. That said, open source tools such as those provided by the creepily named HuggingFace mean I can potentially be creating and refining my own tools rather than relying on third party services. But whether I make or buy I think these uses all help me engage primary sources rather than replacing my engagement.
So readers, is there anywhere you’re experimenting?
Recent Comments