Comment by zahlman
13 hours ago
No, the bias-making machinery is embedded in the machinery, part of the purpose of which is to do a rough kind of statistical analysis via "attention". If for example "Tufa" keeps appearing (n=small, but more than for the other fake tribes) next to terms indicating skill at some task, of course that will be noticed. It will last for as long as that information is in the context window (weights for the current model don't get updated as a result of conversation; that's just not how they work). And of course that can happen from random chance, and of course the LLM has no way to externally verify the extent to which randomness is in play (or the ground-truth probabilities).
The paper makes clear that they used pre-trained, frontier models — in other words, they did not train models on fake data about the fake tribes that would ascribe fake stereotypes to them. There is nothing to suggest that the training data somehow accidentally encoded biases related to fake tribes that the creators of the training data (i.e. ordinary people going about their ordinary Internet lives) somehow accidentally expressed.
There is also nothing to suggest that reading the entire Internet would somehow predispose the reader towards the general idea of being "biased", in the sense that you would have to have in mind to see an actual problem here. But really, the kind of "bias" we're talking about here is really pattern-matching on the available data, which is a big part of what leads people to apply the term "intelligence" to the models. See also the way that people try to make "culturally neutral" IQ tests specifically by having them focus on the ability to infer patterns (e.g. https://en.wikipedia.org/wiki/Raven's_Progressive_Matrices ).
I think you're slightly misunderstanding my point, which is that a huge spectrum of associative pattern seeking logics are embedded in language and that the LLM learns them and operates them, approximately.
"There is nothing to suggest that the training data somehow accidentally encoded biases related to fake tribes that the creators of the training data (i.e. ordinary people going about their ordinary Internet lives) somehow accidentally expressed."
This is exactly not what I'm suggesting.
I think it's less obtuse to make the point that there are social-prejudicial structures encoded in the latent space. "associative pattern seeking logics" is a bit incoherent, and distributional semantics doesn't live at a level accessible to cultural analysis and theoretics, IE film critique.
If you walked into a film theory class and posited that you could derive every single encoded interpretation of a film by memorizing the positioning of the actors and objects frame-by-frame, plus the audio track in another language that you do not speak, for every single piece of video ever made, you'd probably be asked to leave. Not that I'm advocating for the position of critique here, I don't think the anti-distributional semantics crowd is ever going to recover from their humiliation that's been accelerating over the last 8 years. It's just that from the position of critique it requires a coherent narrative that human brains are capable of ingesting (IE not maximal information overload).
If you really want to go the lower level route, I think Francois Laruelle's non-philosophie touches on what you might be thinking of in a much more robust way, shining a light on the unexamined consequences of decision and dialectics of-themselves. If you can stomach the writing of continentals, that is.
> the anti-distributional semantics crowd
Who is this crowd specifically? The stochastic parrots people? Noam Chomsky? I don't think they're good representatives of media theory at all whatsoever. The humanities are much more diverse than they're made out to be in this crap AI culture war.
> distributional semantics doesn't live at a level accessible to cultural analysis and theoretics
They might not have computational access but the theories are all about contextuality, for example Jacque Derrida's "trace" was the first thing that came to mind when I saw this headline. Those people are tuned in on the microscopic level to what LLM researchers are bumping into on macroscopic scales. I'm thinking of post-structuralists especially. But all kinds of people and I'm sure what's happening right now is way more interesting than our crude labels ("the post-structuralists", "the anti-distributional semantics crowd") could actually do justice.
> you'd probably be asked to leave
It would depend a lot on the specific school and instructor but in general I really don't think you would. I've taken classes like this and people were far more open minded and critical than you might assume. And my broader point is that there are really sharp conversations happening in these spaces for decades.
> I think it's less obtuse to make the point that there are social-prejudicial structures encoded in the latent space. "associative pattern seeking logics" is a bit incoherent
Yeah I definitely could be less sloppy but I my point is that language encodes not just word semantics but entire ways of thinking, and they're encoded at multiple levels and in superposition. By "associative logics" I hand-wavingly mean all manner of categorical thinking, "amygdala" thinking, mapping, putting things into buckets, hedging. These kinds of cognitive habits are everywhere in language, and are culturally situated "distributional semantics" style. It doesn't surprise me that when we simulate them with LLMs we'd get results like this because I've studied a little bit of cultural theory in the past and they were on this stuff forever ago.
By the way I actually do think there's more to it than just distributional semantics, but not in any way that would downplay the potency of that theory. Moreso I'm curious about generalizations of the distributional idea into TDA and category theoretic approaches. As well, there's a lot of really cool quantum-like modelling emerging in applied math that I can only see getting more relevant if/when quantum computers come around and quantum models become runnable.
It could be that a model prefers the tribe mentioned in closest proximity to the word candidate most of the time. It could be that it prefers the one that's third in a series. It could be that it prefers the one with even numbers of letters.
The model is biased. That's it's entire function, to bias certain tokens over other tokens based on a bunch of vectors and context. There's no telling what is influencing that bias.
The models will be statistically more likely to choose one of the options for completely unknowable reasons.