Right now there’s a lot of chatter about AI tools, specifically for writing text and generating images.
I’ll collect my takes as of now in this place and we’ll see how well they age.
Takes incoming now.
I had my first brushes with AI/neural net text generators in 2017. I read an article by Janelle Shane about the bizarre and charming paint color names she was able to coax out of an unspecified neural network.
I was enthralled and hooked. I read her other posts and learned a little more, enough to find talktotransformer.com (it redirects now.)
It was a lot of fun because the results were usually surprising in some way. Bizarre and often hilarious jargon would spill out of that thing and it was really fun. I shouldn’t use the past tense here because as of this writing it’s still available to use.
Now there’s a breakthrough AI tool we all can try, chatGPT. It’s obviously worlds more advanced than the neural net at talktotransformer.
It is a great emulation of a very mediocre writer.
Nassim Taleb wrote in Fooled By Randomness 20+ years ago:
Alan Turing came up with the following test: A computer can be said to be intelligent if it can (on average) fool a human into mistaking it for another human. The converse should be true. A human can be said to be unintelligent if we can replicate his speech by a computer, which we know is unintelligent, and fool a human into believing that it was written by a human.Nassim Taleb, Fooled by Randomness
Note: I’m not sharing examples of AI generated text because I’m finding it to be really boring when others do it.
So at least for now we can use chatGPT as a ruler to judge competent copy by.
Nicole Hemsoth outlines some really important concerns about the sheer amount of mediocre and factually questionable copy this technology will inevitably create in this very good article.
We go back again to the theme of math and volume and such to address the most important point: the information danger is an exponential problem. One series of mistakes generated, then repeated by content mills for a decade, means those problems get trained into the core AI language model from the corpus of the internet and reinforced.Nicole Hemsoth https://www.theregister.com/2022/12/06/internet_ai_gpt_ios/
These AI tools are introducing us to new problems of scale.
Nobody begrudges an artist his or her influences as long as their work is personal and honest (to the degree that kind of thing is apparent in a painting or some such.)
Now train an AI image tool on hundreds of an obscure artist’s works to create unlimited stylistically identical works and it seems like that artist is perhaps being taken advantage of.
I’ll update in the next day or so and talk about these image tools.
edit 12/9/22 chat GPT answers questions like a beauty pageant contestant or political press secretary.