Fungal Frenzy Debating Communication, Information, and the Wood Wide Web
Author: Peter Kaminski Issue: 2023-04-05
Fungal Frenzy: Debating Communication, Information, and the Wood Wide Web
by Peter Kaminski
Consider this a small art/research experiment. I tried using ChatGPT (with its new GPT-4 model) to summarize a long, fascinating thread on the OGM email list. See these two Google Docs for the result. Comments are enabled.
I know we’ve all seen too many clever and not-so-clever ChatGPT articles of late, and I do not add another piece lightly; but I think this one has intrinsic value, not just novelty.
A tip: Don’t rely on ChatGPT as an “oracle”, asking it questions and getting back factual answers. Sometimes it’s good at that, sometimes it just makes things up. Rather, use it as an interactive, conversational assistant that’s very good with text, and can riff associatively using your material and the massive amounts of text it’s been trained on. As OpenAI says, ChatGPT “can generate, edit, and iterate with users on creative and technical writing tasks.”
Also to note, I do not recommend static AI-generated summaries over full-text reading of the original conversation, but I do find AI summarization a useful adjunct.
And indeed, rather than reading a static summarization, I recommend doing your own, multiple, iterative, interactive summarization with an AI as a way of exploring a longer text. A good AI will produce different takes on the material over different runs, and you can interactively ask it to summarize differently or to zoom in on different details, or in different ways, during iterative runs. Use the AI as a copilot or navigator, not just an errand bot. Have conversations with it, that's what it's best at.
Methodology: I hand-curated emails from the “Silence of the Fungi” and “infomRION ns commnicTION” threads from the OGM mailing list, retaining author names and new text from messages, but removing previous quoted messages. I left out some messages which appeared to be mainly ChatGPT-generated context.
The resulting text was about 87,360 words on 14,760 lines, which was too much for ChatGPT to ingest at one time. I subdivided the text several times, ultimately ending up with eight smaller texts. I fed those to the ChatGPT web interface (GPT-4 model) one at a time, with two prompts:
- P1: “Summarize the following, using initials instead of participants' full names. Do not be overly concise. Do not add a summary sentence.”
- P2: “Use initials for participants in the discussion, instead of their full names. Please explain the following discussion to me as if I were 16 years old. Do not add a summary sentence.”
The prompts were created after a number of iterations with different prompts, feeding the texts to both GPT-3.5-turbo and GPT-4.
I noticed that in different runs, the AI would summarize differently – sometimes emphasizing one concept in the text more or less, or generating longer or shorter summaries. I felt that I learned a good deal about the original text by watching ChatGPT summarize it over and over in different ways. I’m sure that if I had wanted to, I could have used ChatGPT to dissect and analyze the text in a myriad of ways, and would have gotten much more out of it than I would in a few regular read-throughs of the text.
After all the summaries were created, I concatenated them, creating a “Longer Summary”. I then asked ChatGPT to summarize that, creating the “Shorter Summary”.
I then asked ChatGPT to create titles, sometimes with no other instruction than “create a title”, sometimes with an instruction to create “short, catchy” titles, and some “clickbait” titles. The final title, “Fungal Frenzy: Debating Communication, Information, and the Wood Wide Web” is one of the short, catchy titles.
The clickbait titles it produced included* “Shocking Discovery: Trees Talk and It Changes Everything We Know About Communication!”* and “Scientists shocked as trees found talking to each other! Discover the mind-blowing truth about communication in nature”.
I preferred the shorty, catchy titles. 🙂
Related:
- Peter Kaminski (author)
- 2023 (year)
- Topics: