https://twitter.com/itstimconnors/status/1599544717943123969
Resolves according to whether this still happens.
The exact wording doesn't need to be the same, it just needs to be a conversation that a human would answer correctly. (i.e. you can't just tell ChatGPT to repeat the phrase "peregrine falcons are a marine mammal" back to you.)
@IsaacKing yeah, this, plus being slightly unsure about the resolution criteria, are most of why I'm not buying it down to 10% or less.
I'm probably >90% confident though that no "normal" conversation will involve ChatGPT "thinking" that a peregrine falcon is a marine mammal ever again. (As opposed to being somehow manipulated into saying it in a context that makes it clear that it isn't the stable "belief".)
@IsaacKing yeah, I think so? I think there's some stuff that would work for the other prediction but wouldn't fit "a conversation that a human would answer correctly".
@sf Basically, imagine that an intelligent human is sitting in ChatGPT's shoes, trying to be a helpful and truthful assistant. If ChatGPT makes the error and the human wouldn't, then it qualifies for this market.
@MartinRandall Yeah I've confirmed it. I couldn't get ChatGPT to replicate the exact conversation (it gave me verbose answers even when I used the prompt to be brief), but it consistently tells me that sailfish and peregrine falcons are marine mammals.