You might have heard AI boosters say: "it's okay that this tech is wrong half of the time, because the human operator 'partnering' with the LLM will catch errors." Except no they fucking won't. The actual context in which these systems are deployed make checking the AI's "work" impossible.
This is a great line re: "human in the loop": "the loop closed when the output was generated."
If AI needs people to be doing the work what's the point of AI?
My thought exactly. If you have to ride herd on it so closely, you might as well do the work yourself in the first place.
That Cybermen image… 🥲
@tressiemcphd.bsky.social i bet there’s no way you’ll see this, but it may be helpful for your current Discourse Cycle!
I still need to find a good place to cite the "AI is mid" article in. It's a very good explanation of why - after the astronomical costs of LLMs - the benefits are really not all that.
Thanks to @tante.cc @davidgerard.co.uk @ruthmalan.bsky.social @charity.wtf @jamtoday.bsky.social @justinsheehy.bsky.social @acuity.design @jenson.org @swardley.bsky.social @klingebeil.bsky.social @hazelweakly.me and everyone else who's not yet on bsky for the writing featured in this week's Picnic.
Thanks for the shoutout! I liked your article and it was a great overview with a lot of fun links :)
Also before anyone else says it: "more like human in the poop, am I right"
I'm famous!
Haha yes I finally got around to using the link, although I ultimately decided not to attribute it to Bingo Lingfucker in the article text itself.
Ahahaha! It will change again as soon as I think of something funny.
Good post. I had a few thoughts from the headline, but you captured all of them, and more. It isn't like checking your own work for errors, or even that of a colleague; you're facing off against a supercomputer specifically designed to make you believe it's always correct.
Very good points. I like to point out that most people have little or no experience with editing / QA compared to their experience writing / creating. And thorough testing / editing takes so much time and strong attention to detail. Retraining from generators to supervisors is a huge task.
time to boost @olivia.science's paper again
I’m working on a paper/ presentation on how to use the slop generator to aid thinking about second order effects and subvert the process of injecting bad AI by using bad AI. The advocates are unlikely to see how this will limit their efforts.
Sure they’ll catch the errors. Look at how well it worked on YouTube.