Follow

here's a recording of big ol' lecture I wrote & delivered for BMOLab and Vector Institute last month, entitled "Language Models and Poetics," in which I claim that computers not only *can* generate poetry, but in fact they can *only* generate poetry.

bmolab.artsci.utoronto.ca/?p=8

the discussion touches on GPT-n (of course), speech act theory, William Carlos Williams' _Spring and All_, Frank Lantz (twice), and more 🎶

(the audio didn't come out great, happy to supply text/slides to interested folks)

(also I did not pick a great camera angle hahaha whoops)

Show thread

@aparrish just watched this and it's great! am wondering if there's a distinction between speech acts generated by a statistical model, and those that arise from something more simulation-based. is it really empty and void when dwarf fortress tells me my dwarf has died? that's a fact about the (a?) world, not a mere citation

@mewo2 there is a really great chapter (chapter 3) about this very thing in James Ryan's PhD dissertation, which even references speech act theory, and I have no idea why I didn't think to discuss it in the talk! I'm not sure how to answer. On the one hand, it's a question about fiction/nonfiction & simulated events, but on the other it's a more basic question of artifacts that relay information, like—is it empty and void when my car's speedometer shows me how fast I'm going?

@mewo2 (the extreme version of my argument would claim that the speedometer doesn't state anything, that in fact the car's manufacturer etc. does the stating, and there's truth to that, bc if the speedometer is wrong, it's not the car's "fault." following this to its logical conclusion, the output of language models *isn't* hollow & void, but counts as a sort of speech act on the part of those that created it, i.e. "the following is a sequence of tokens that have certain statistical properties")

@aparrish yeah, I think I'm less interested in the fiction/non-fiction distinction than the second question - it feels to me like purely statistical manipulation of language tokens is qualitatively different to language being used as a way of relaying information, even if it's coming out of a machine. your argument makes perfect sense in terms of language models, but those aren't the only way that computers generate language (even if they are the flavour of the month)

@mewo2 yeah, you may be right. on the one hand I want this argument to apply across all techniques, but on the other hand I wish I'd thought of this because then I'd have had a way to make the talk be more specific to language models, haha

@aparrish yeah, I think if you take this too far you end up in chinese room territory, and it definitely feels like humans and speedometers are not doing the same thing, but the distinction is not quite the same as the one you're drawing

Sign in to participate in the conversation
Friend Camp

Hometown is adapted from Mastodon, a decentralized social network with no ads, no corporate surveillance, and ethical design.