emily bender gave a whole talk about how LLMs are not a valid subject of study for computational linguists - ThreadSky

theophite.bsky.social • 86 days ago

emily bender gave a whole talk about how LLMs are not a valid subject of study for computational linguists

Are linguists paying a lot of attention to LLMs? Because this seems like a fascinating finding with large implications: LLMs share highly abstract grammatical concept representations, even across unrelated languages, so even models trained mostly on English do well in other languages.

Comments

leonderczynski.bsky.social•85 days ago

which talk??

keytryer.bsky.social•86 days ago

Linguists today aren't very happy with evidence of blank slatism

theophite.bsky.social•86 days ago

i am a mild LLM skeptic, but i have a lot of disdain for people who do not consider them interesting as a mathematical object which might tell us things about the nature of language

niall-mcauley.bsky.social•86 days ago

Meanwhile, AI is stealing everyone's work and supercharging global warming.

sashotodorov.bsky.social•86 days ago

In a vacuum they are fascinating, but it gets hard to care about them when people are declaring that they found the spark of God in Copyright Infringement Clippy.

parkerkosticak.bsky.social•86 days ago

The crime was showing LLMs to investors. It was cool enough to scam them into building damn nuclear plants.

0kultra.bsky.social•85 days ago

Nuclear plants are good and if LLMs trigger a nuclear buildout and then crash the market in a bubble it would still be worth it because we'll still get nuclear plants

chrononaut.bsky.social•86 days ago

The most uncool and scammy people are overhyping AI and creating a backlash against the technology that never existed before.

IP theft and mass layoffs "because AI will do it all now" also did not help.

I guess for less online people it's easier to just evaluate the tech on its merits.

hrtraulsen.bsky.social•86 days ago

As a know-er of chomsky from not linguistics who thus learned a little bit about linguistics: my understanding was the field kind of moved away from him but does this have the potential to validate his theoretical program a bit?

theophite.bsky.social•86 days ago

i am probably not exactly the person to talk to about this because i am extremely negative about Chomsky, but my general feeling is that his theoretical program as it stands has been cut back to the point where it's a trivial thesis which is trivial to validate, and grounded in the wrong field

hrtraulsen.bsky.social•86 days ago

ty!! ling was never a side interest I ever engaged in super heavily and appreciate a little hopping off point from my odd assortment of half-remembered bits about the field

segyges.bsky.social•86 days ago

it amuses me how much ground "mild llm skeptic" covers in our current environment

conjurial.bsky.social•86 days ago

everything between Gary Marcus and "the singularity is LITRALLY happening by Q3", not inclusive

segyges.bsky.social•86 days ago

Why not Q2? I'm sure there's some of those around

zamfir.bsky.social•86 days ago

Q2 means you're not a skeptic. Sketics think the rapture will come before q4, possibly even later.

gothedistance49.bsky.social•86 days ago

Agreed. As a linguistics major in college I think LLMs are a lot more interesting than some of the stuff Chomsky came up with...

groda.eu•85 days ago

A lot of the LLM hate comes from them being pushed on people who have not asked for them and do not want them. If people *force* you to eat vanilla ice cream you may develop irrational hate towards vanilla ice cream

mrkeithtaylor.bsky.social•85 days ago

You may be right, but my disdain for a thing built using my stolen copyrighted works overwhelms any interest I may have in it. It's like being pestered to appreciate the beauty of the fire as it burns down your house.

davedun.bsky.social•86 days ago

I’m a language teacher, not a linguistic, and am also an “AI” skeptic. My issues are with the hype (we don’t need LLMs *everywhere*) and the ethics (scraping, permission, copyright, energy, water, etc). Could they be if interest? Surely. But too many cons and too few pros rn.

nirebryce.breadthcharge.net•73 days ago

hey so, I was thinking about this paper again today after reading this stackoverflow answer about bluetooth encryption, and the author talks about language-knowledge attacks (frequency analysis ofc among others) on audio encryption

https://security.stackexchange.com/a/262579

nirebryce.breadthcharge.net•73 days ago

and it made me wonder if there's a way to exploit the fact that it does language-solving with similar ways for each language to like,

"feed the machine sentence-shaped encrypted audio garbage and have it spit out an English document" as long as you can split the data into 'word-frames'

nirebryce.breadthcharge.net•73 days ago

anyway i have a brand new threat model now

jordantcarlson.bsky.social•86 days ago

sincerely: most of the people I know who find them useful find them useful mainly for coding automation and time-saving on tasks they know well. what other use cases do you see real potential in? I have an instinctive aversion due to the (very obvious) hype machine hucksters, but I'm curious.

paulboldrin.bsky.social•85 days ago

How are you going to tell your robot butler what to do without an LLM? Not entirely joking.

theophite.bsky.social•86 days ago

notebookLM is very good

jordantcarlson.bsky.social•86 days ago

Was not aware of it; skimming their page, that's the exact use case I would be most skeptical of it based on seeing performance of others on summarising. May have to make time to test it vs. things I already know (and have done manually). No local instance option, though?

kelsig.on.computer•86 days ago

is this just downstream from the ai moral panic or do linguists have justification here

segyges.bsky.social•86 days ago

it's philosophical dualism hiding behind other terminology, moral panic, and the fact that LLMs support a basically statistical or complexity-theoretic theory of language and mind, which is not really amenable to the toolkit or theories that linguists traditionally truck in

cinera-verinia.bsky.social•85 days ago

They should learn statistics and complexity theory then.

jephalopod.bsky.social•86 days ago

I’m pretty sure the part of my brain that gets me into trouble is some kind of LLM

struthious.bsky.social•86 days ago

I mostly don’t like how they make arguments based on far-from-consensus positions in philosophy of language without acknowledging that those assumptions are disputed