Enjoying this editorial cartoon & commentary by Phie Wei in the Scot Scoop regarding my AI lore books. Wei writes:
Author Tim Boucher claims to have written 97 books in nine months using ChatGPT. A sense of admiration exists for someone who has taken advantage of modern technology to create content at unimaginable speeds. However, seeing creativity being removed from the writing process is concerning. ChatGPT’s ability to use information derived from formerly published information online falls short because it can’t say or bring anything new to the world. Taking words from profound works of art without crediting the author and retooling them in a new context devalues authors’ dedication to their writing.
I wrote asking if they’d like to do an interview, but to give a brief reply here:
- Creativity isn’t being removed from this process. The locus of activity of the artist is simply shifted & the writer becomes more like a creative director or a film director. In short, I’ve spent last 9-10 months using AI tools in a kind of feverish deeply creative and deeply intuitive “fugue state” that rivals or exceeds any other creative endeavor I’ve ever engaged in.
- I don’t think LLMs or image diffusion models having been trained on other source works means that they can’t bring anything new into the world. First, because they are able to uncover connections that no humans have ever conceived of before, and to sift through those connections for the “best” ones (however we define that) in record speed. Simply look at how AI models are being used in pharmaceutical research for one example among many. All of that is new information being brought into the world. Second, its the job of the artist/writer as creative director to make certain any creative works produced DO bring something new into the world!
- LLMs do not take words from other works, as I wrote about a bit here. To recap, AI models do not “sample” or take specific excerpts from other works, as in a collage. What they do is take measurements of dimensionality (e.g., they study and compare attributes) of a vast corpus of training material, and use that to uncover statistical semantic relationships – webs of meaning or common connections. Thus, there is really no single work which AI generated content is specifically referencing (unless you specify it – like by prompting, “A cat in the style of the Mona Lisa”).
There’s more to say of course, but that’s as good a foundation to go from as any.