Re:Misalignment and Deception by an autonomous stock trading...
by og_kalu on Monday November 20, 2023 @05:27PM
(#38355798)
Replace GPT-4 with Human in this paper and nearly all the found behaviour is entirely predictable and even expected.
Now replace the behaviour with "weird bugs in the computer program" and it's mostly just confusing, unintuitive and unpredictable. Traditional bugs don't even work this way.
The conclusion here is not necessarily that GPT-4 is a human but that ascribing human like behavior results in a powerful predictive model of what you can expect in GPT-4 output.
What would be silly is throwing away this predictive model for something worse in every way just so I can feel more special about being human.
A statistical model of language patterns is not a person. A statistical model of language patterns is not a living thing. Please go touch grass.