Since LLMs are by definition based 100 percent on what is written down, they are gloriously ignorant of everything about the human condition that isn't written down.
Suppose an LLM were asked, “Will people like New Coke or Old Coke better?” If an LLM were conscious of its limitations, it would answer, "How on earth would I know?”
The executives in charge of that real-world decision would probably be similarly inclined to come up with an enthusiastic opinion. "conscious of its limitations" is a rare trait among (male?) humans :-) More, it's the impossibility of an LLM saying "f*ck off, mate" that tells me they'll never adequately simulate the animal essence of being alive.
Interesting to tease out the differences btwn the practice of description and prediction. But with regard to Weizenbaum: would he have cared very much about these distinctions? For him, the important diff was btwn the practice of instrumental reasoning and moral reasoning. Those map imperfectly to the differences btwn otoh the practice of *describing* the present or the future and otoh *prescribing* what the present and future ought to be. Still, those similarities and differences are also worth teasing out and relating back to Weizenbaum.
It is wild, but also perfectly compatible with the last year or two, that many are bullish that that silicon people can reproduce some of the measurable parts of response distributions of meat people. Have you seen any papers yet measuring treatment responses from chatbot therapy vs. more traditional talk therapy, or the deployment of LLMs instead of "experts" in other fields? Replacing "experts" seems more difficult than pretending that LLMs can represent samples for social science research. Related, it seems important to know whether LLMs can be induced to reproduce the full distribution of so-called personality tests. I'd wager that the incidence of obsequious narcissism is quite a bit higher than in the general meat-population.. but who knows!
Since LLMs are by definition based 100 percent on what is written down, they are gloriously ignorant of everything about the human condition that isn't written down.
Suppose an LLM were asked, “Will people like New Coke or Old Coke better?” If an LLM were conscious of its limitations, it would answer, "How on earth would I know?”
The executives in charge of that real-world decision would probably be similarly inclined to come up with an enthusiastic opinion. "conscious of its limitations" is a rare trait among (male?) humans :-) More, it's the impossibility of an LLM saying "f*ck off, mate" that tells me they'll never adequately simulate the animal essence of being alive.
the fake respondents at least answer the survey, which is more than i do
Interesting to tease out the differences btwn the practice of description and prediction. But with regard to Weizenbaum: would he have cared very much about these distinctions? For him, the important diff was btwn the practice of instrumental reasoning and moral reasoning. Those map imperfectly to the differences btwn otoh the practice of *describing* the present or the future and otoh *prescribing* what the present and future ought to be. Still, those similarities and differences are also worth teasing out and relating back to Weizenbaum.
It is wild, but also perfectly compatible with the last year or two, that many are bullish that that silicon people can reproduce some of the measurable parts of response distributions of meat people. Have you seen any papers yet measuring treatment responses from chatbot therapy vs. more traditional talk therapy, or the deployment of LLMs instead of "experts" in other fields? Replacing "experts" seems more difficult than pretending that LLMs can represent samples for social science research. Related, it seems important to know whether LLMs can be induced to reproduce the full distribution of so-called personality tests. I'd wager that the incidence of obsequious narcissism is quite a bit higher than in the general meat-population.. but who knows!
I'm really curious what happens to the distribution of fake sampling. There aren't any tails in such a sample.