Steve Olson

I've been listening and reading on AI, and it strikes me how most of the real questions around artificial intelligence are far from new, and not technological at all.

So I laughed really hard at this story of a dog who, in 1908, was rewarded with steak for saving a child who'd fallen into the Seine. Naturally, he started pushing kids in to rescue them to get more steak. Much like a bot optimizes to its parameters. (the fancy term is "principal-agent" problem). https://www.nytimes.com/1908/02/02/archives/dog-a-fake-hero-pushes-children-into-the-seine-to-rescue-them-and.html

DOG A FAKE HERO.; Pushes Children Into the Seine to Rescue Them and Win Beefsteaks.

The New York Times
Jan 18, 2023 at 8:58pmWeb
Show threadJesseJan 18, 2023
@steveolson THE K9 ALIGNMENT PROBLEM
Show threadSteve OlsonJan 18, 2023
@misc RIGHT?!
Show threadJesseJan 18, 2023
@steveolson Rick and Morty tried to warn us.
Show threadJef PoskanzerJan 19, 2023
@steveolson See also the first half of "The Good, the Bad, and the Ugly".
Show threadSteve OlsonJan 19, 2023
@jef hahhahahahhaha well played
Show threadDr Zoyander StreetJan 19, 2023
@steveolson I honestly think that we intentionally set out to train computersl the way we train animals. Something something history of technology starts with companion species something something Donna Haraway.
Show threadbluestarultorJan 20, 2023
@steveolson "Anyone who thinks human-like A.I. is impossible needs to understand that it's already got laziness down pat" seems to be an appropriate takeaway. XD
Show threadSteve OlsonJan 20, 2023
@bluestarultor hahahahahahaha. fair point.
Show threadkithJan 20, 2023
@steveolson The podcast CRIMINAL featured this story last year; it is _hilarious_.
Show threadAlerta! Alerta!Jan 20, 2023

@steveolson Which is not only an issue with AI or dogs. It is a general issue with any metric that is used as an incentive.

Whether that is the function between rescued kids and steaks or solved issues and salary...

Terry Pratchett described that nicely in one episode where the city paid a dollar per rat-tail to get a grip on the rat plague. Some time later they started taxing rat-farms....

Show threadSteve OlsonJan 20, 2023
@heiglandreas heh. yeah, i like I said, that's what kind of dawned on me... that the new-fangled AI questions really are just age-old problems we have to apply in a new context, albeit one that could potentially kill us all in an attempt to meet its parameters.
Show threadGeoff 🏴󠁧󠁢󠁳󠁣󠁴󠁿Jan 20, 2023
@steveolson Literal AI version: it was trained on Tetris with "play for as long as possible" as the goal. The AI found the pause button.
For human analogues see most episodes of #Taskmaster.
Show threadYohan John 🤖🧠Jan 20, 2023

@steveolson

Nice! This is also related to Goodhart's Law.

Show threadSteve OlsonJan 20, 2023
@DrYohanJohn heh, i’d forgotten about that one!