The distinction between what might be called “lying” and “bullshitting” is important here, because they scale with competence differently.
It was pretty interesting watching this develop in my kids. Saying “No!” to “Did you take the cookie from the cookie jar?” is the first thing you get, because it doesn’t require a concept for “truth” at all. Which utterance postpones trouble? Those are the sounds I shall make!
Yet for a while my wife and I were in a situation where we could just ask our younger kid about her fight with our older kid, because the younger kid did not have a concept for fabricating a story in order to mislead. She was developed enough to say “No!” to things she knew she did, but not developed enough to form the intention of misleading.
The impression I get from LLM is that they’re bullshitters. Predict what text comes next. Reward the one’s that sound good to some dumb human, and what’s gonna come out? We don’t need to postulate an intent to mislead, we just need to notice that there is no robust intent to maintain honest calibration—which is hard. Much harder than “output text that sounds like knowing the answer”.
It takes “growing up” to not default to bullshit out of incompetence. Whether we teach them they’ll be rewarded by developing skillful and honest calibration, or by intentionally blowing smoke up our ass is another question.
The distinction between what might be called “lying” and “bullshitting” is important here, because they scale with competence differently.
It was pretty interesting watching this develop in my kids. Saying “No!” to “Did you take the cookie from the cookie jar?” is the first thing you get, because it doesn’t require a concept for “truth” at all. Which utterance postpones trouble? Those are the sounds I shall make!
Yet for a while my wife and I were in a situation where we could just ask our younger kid about her fight with our older kid, because the younger kid did not have a concept for fabricating a story in order to mislead. She was developed enough to say “No!” to things she knew she did, but not developed enough to form the intention of misleading.
The impression I get from LLM is that they’re bullshitters. Predict what text comes next. Reward the one’s that sound good to some dumb human, and what’s gonna come out? We don’t need to postulate an intent to mislead, we just need to notice that there is no robust intent to maintain honest calibration—which is hard. Much harder than “output text that sounds like knowing the answer”.
It takes “growing up” to not default to bullshit out of incompetence. Whether we teach them they’ll be rewarded by developing skillful and honest calibration, or by intentionally blowing smoke up our ass is another question.