I appreciate the people who are doing empirical research that demonstrates the problems with this sort of thing. What I’d really love to see is a #philsci take on the question. What is the epistemic value of an LLM simulated research subject? Has anyone written anything on this? - ThreadSky

awmercer.bsky.social • 319 days ago

I appreciate the people who are doing empirical research that demonstrates the problems with this sort of thing.

What I’d really love to see is a #philsci take on the question. What is the epistemic value of an LLM simulated research subject?

Has anyone written anything on this?

Reposted from CREST Sociology

NEW PREPRINT by CREST Sociology's @scoavoux.bsky.social and @ppraeg.bsky.social: "Machine Bias. Generative Large Language Models Have a Worldview of Their Own"

LINK: doi.org/10.31235/osf...

Comments

recombobulating.bsky.social•319 days ago

The epistemic value seems to largely be control and predictability. That they are seen as easier and cheaper to deal with than human beings.

recombobulating.bsky.social•319 days ago

This paper seems to be attempting to locate the bias in the machine rather than in the human generated data it is built and trained on and I find that rather odd.

svateboje.bsky.social•319 days ago

if we distinguish btwn the transformer approach as a form for working with sequential data & what the sequential data are taken from, we can think about how transformers match up to the correlational structure of rather short instances of language. not much philsci, but characterizes the simulations

svateboje.bsky.social•319 days ago

another angle is the unity of the data. there is one guess at the next term which is conditional on the entire corpus. it is hard to imagine what odd analogues to simpson’s paradoxes might be out there. but, surveys are ofc centrally concerned w breaking that kind of unity up.

svateboje.bsky.social•319 days ago

ie sticking w language data, where is there any thing like weighting based on prior knowledge of population structure? just nothing at all happening like that. or, how close can the hermeneutical task of question design actually get to the novel puzzles of prompt design?

awmercer.bsky.social•319 days ago

Yeah. I guess you can think of an LLM in this context as giving you Pr(Y|X) for an extremely high dimensional and complicated function of X. To go from there to inferences about Y, you need to know Pr(X) in the population.

awmercer.bsky.social•319 days ago

But I don’t think one could really even articulate what X is when you’re talking about a corpus of text data. Or if all values of X in the population are even present within the training data. And that’s not even taking into account questions of algorithmic bias.

awmercer.bsky.social•319 days ago

Or maybe it’s that if you could articulate what X is in the context of the training data, I’m not sure how you could get to Pr(X) in the population of interest which exists outside of and prior to the training data.

Comments

Posting Rules

Reply