Idiot AI

Logo, ChatGPT by OpenAI

AI would be artificial intelligence, aka ChatGPT. I've read a couple of science fiction novels by Peter Watts, and in some of them our hero is fighting against an idiot AI. The techniques the hero uses to get around the machine brain are interesting. I haven't got them entirely sorted, but it makes for an entertaining stories.

Hand Waving Freakoutery has a post up about ChapGPT and the biases that are built into it. It's kind of long, but you don't really need to read the whole thing, the first example is enough to give you the general idea. His conclusion is what I enjoyed:

While it’s certainly fun to trash the Woke Egregore with this sort of exercise, I daresay a LLM [Large Language Model] trained against the MAGA [Make America Great Again, i.e. Donald Trump] echo chamber could probably be just as easily trolled to produce just as absurd statements. But I’m bothered that the people arguing in this AI space today are arguing either against the idea that ChatGPT is biased, or arguing against making new biased tools, when the most useful thing in my mind would be to build biased ones on purpose.

With an echo-chamber specific LLM, you can talk to the Egregore directly. You can find out what it thinks, what it wants, and what it needs to propagate itself. You can use that dialogue for marketing reasons, to sell more MAGA hats on your pillow website or more fake Antifa membership cards. Roger Goodell could train a LLM against every post on every NFL Football forum on the internet, and then ask OblongSpheroidGPT whether he should make changes to the late-hit rules. The tech nerds could put focus groups of all kinds out of business, simply by training LLMs to enact the will of the Penultimate Soccer Mom Egregore.

And at the pinnacle of it all, we could LLM the major culture war egregores themselves and just have them fight each other, flooding Twitter with million man armies of bots to wage wars while we go outside and throw that pigskin around ourselves.