An anthropomorphic model of the software, wherein you can articulate things like “the software is making up packages”, or “the software mistakenly thinks these packages ought to exist”, is the right level of abstraction for usefully reasoning about software like this. Using that model, you can make predictions about what will happen when you run the software, and you can take actions that will lead to the outcomes you want occurring more often when you run the software.
If you try to explain what is going on without these concepts, you’re left saying something like “the wrong token is being sampled because the probability of the right one is too low because of several thousand neural network weights being slightly off of where they would have to be to make the right one come out consistently”. Which is true, but not useful.
The anthropomorphic approach suggests stuff like “yell at the software in all caps to only use python packages that really exist”, and that sort of approach has been found to be effective in practice.
in terms of communication utility, it’s also a very accurate term.
when WE hallucinate, it’s because our internal predictive models are flying off the rails filling in the blanks based on assumptions rather than referencing concrete sensory information and generating results that conflict with reality.
when AIs hallucinate, it’s due to its predictive model generating results that do not align with reality because it instead flew off the rails presuming what was calculated to be likely to exist rather than referencing positively certain information.
it’s the same song, but played on a different instrument.
Can we fucking stop anthropomorphising software?
No?
An anthropomorphic model of the software, wherein you can articulate things like “the software is making up packages”, or “the software mistakenly thinks these packages ought to exist”, is the right level of abstraction for usefully reasoning about software like this. Using that model, you can make predictions about what will happen when you run the software, and you can take actions that will lead to the outcomes you want occurring more often when you run the software.
If you try to explain what is going on without these concepts, you’re left saying something like “the wrong token is being sampled because the probability of the right one is too low because of several thousand neural network weights being slightly off of where they would have to be to make the right one come out consistently”. Which is true, but not useful.
The anthropomorphic approach suggests stuff like “yell at the software in all caps to only use python packages that really exist”, and that sort of approach has been found to be effective in practice.
“Hallucinate” is the standard term used to explain the GenAI models coming up with untrue statements
in terms of communication utility, it’s also a very accurate term.
when WE hallucinate, it’s because our internal predictive models are flying off the rails filling in the blanks based on assumptions rather than referencing concrete sensory information and generating results that conflict with reality.
when AIs hallucinate, it’s due to its predictive model generating results that do not align with reality because it instead flew off the rails presuming what was calculated to be likely to exist rather than referencing positively certain information.
it’s the same song, but played on a different instrument.
What standard is that? I’d like a reference.