I don’t know if it’s just my age/experience or some kind of innate “horse sense” But I tend to do alright with detecting shit responses, whether they be human trolls or an LLM that is lying through its virtual teeth. I don’t see that as bad news, I see it as understanding the limitations of the system. Perhaps with a reasonable prompt an LLM can be more honest about when it’s hallucinating?
I don’t know if it’s just my age/experience or some kind of innate “horse sense” But I tend to do alright with detecting shit responses, whether they be human trolls or an LLM that is lying through its virtual teeth
I’m not sure how you would do that if you are asking about something you don’t have expertise in yet, as it takes the exact same authoritative tone no matter whether the information is real.
Perhaps with a reasonable prompt an LLM can be more honest about when it’s hallucinating?
So far, research suggests this is not possible (unsurprisingly, given the nature of LLMs). Introspective outputs, such as certainty or justifications for decisions, do not map closely to the LLM’s actual internal state.
I’m not sure how you would do that if you are asking about something you don’t have expertise in yet, as it takes the exact same authoritative tone no matter whether the information is real.
I agree – That’s why I’m chalking it up to some kind of healthy sense of skepticism when it comes to trusting authoritative-sounding answers by themselves. e.g. “ok that sounds plausible, let’s see if we can find supporting information on this answer elsewhere or, maybe ask the same question a different way to see if the new answer(s) seem to line up.”
So far, research suggests this is not possible (unsurprisingly, given the nature of LLMs). Introspective outputs, such as certainty or justifications for decisions, do not map closely to the LLM’s actual internal state.
Interesting – I still see them largely as black boxes so reading about how people smarter than me describe the processes is fascinating.
let’s see if we can find supporting information on this answer elsewhere or, maybe ask the same question a different way to see if the new answer(s) seem to line up
Yeah, that’s probably the best way to go about it, but still requires some foundational knowledge on your part. For example, in a recent study I worked on we found that programming students struggle hard when the LLM output is wrong and they don’t know enough to understand why. They then tend to trust the LLM anyways and end up prompting variations of the same thing over and over again to no avail. Other studies similarly found that while good students can work faster with AI, many others are actually worse off due to being misled.
I still see them largely as black boxes
The crazy part is that they are, even for the researchers that came up with them. Sure we can understand how the data flows from input to output, but realistically not a single person in the world could look at all of the weights in an LLM and tell you what it has learned. Basically everything we know about their capabilities on tasks is based on just trying it out and seeing how well it works. Hell, even “prompt engineers” are making a lot of their decisions based on vibes only.
I don’t know if it’s just my age/experience or some kind of innate “horse sense” But I tend to do alright with detecting shit responses, whether they be human trolls or an LLM that is lying through its virtual teeth. I don’t see that as bad news, I see it as understanding the limitations of the system. Perhaps with a reasonable prompt an LLM can be more honest about when it’s hallucinating?
I’m not sure how you would do that if you are asking about something you don’t have expertise in yet, as it takes the exact same authoritative tone no matter whether the information is real.
So far, research suggests this is not possible (unsurprisingly, given the nature of LLMs). Introspective outputs, such as certainty or justifications for decisions, do not map closely to the LLM’s actual internal state.
I agree – That’s why I’m chalking it up to some kind of healthy sense of skepticism when it comes to trusting authoritative-sounding answers by themselves. e.g. “ok that sounds plausible, let’s see if we can find supporting information on this answer elsewhere or, maybe ask the same question a different way to see if the new answer(s) seem to line up.”
Interesting – I still see them largely as black boxes so reading about how people smarter than me describe the processes is fascinating.
Yeah, that’s probably the best way to go about it, but still requires some foundational knowledge on your part. For example, in a recent study I worked on we found that programming students struggle hard when the LLM output is wrong and they don’t know enough to understand why. They then tend to trust the LLM anyways and end up prompting variations of the same thing over and over again to no avail. Other studies similarly found that while good students can work faster with AI, many others are actually worse off due to being misled.
The crazy part is that they are, even for the researchers that came up with them. Sure we can understand how the data flows from input to output, but realistically not a single person in the world could look at all of the weights in an LLM and tell you what it has learned. Basically everything we know about their capabilities on tasks is based on just trying it out and seeing how well it works. Hell, even “prompt engineers” are making a lot of their decisions based on vibes only.