GlyphFeb 20
FWIW I really disagree with Cory Doctorow on the "purity culture" thing but I'm not going to get mad about it. The man's job is having takes. He has like fifty takes a day. He is Takes Georg. The vast majority of the takes are fine, and many are actually pretty great. If I had to have that many takes that fast I would have _way_ more terrible takes. I hope he changes his mind on this one but even if not it's fine I don't have to agree with everyone on everything
Show threadR3yScaleFeb 21

@glyph
Firstly, the whole purity culture thing is odd and I think the way Cory has framed the original post as well as his responses around that has definitely soured the entire discourse.

I don't really know how to think through this issue yet but I do appreciate reading your thoughts on the matter. You point toward a more comprehensive criticism of LLM usage which would makes more sense to me, as opposed to several people pointing out that spellcheck already exists. The thing is, Cory explicitly said that in his experience the LLM spellechecker works better.

Show threadR3yScale

@glyph
Of course LLMs have certain biases and hallucinations, but the pre-existing tech also has certain patterns of distrortions and false positives (Cory claims improvement here, that the LLM produces less false positives). The question seems to me to be if those kinds of problems inherent to LLMs, that is things like the tendancy to 'hallucinate' and the unreliability/undeterministic nature, do in fact justify the position that LLM use is fruitless or even harmful in those applications.

I think I do lean toward the position that there is something fundamental about LLM tech that applies in a locally hosted open-source context that does in fact cause problems, there's something about ceding human agency in such a way that disrupts our congitive abilties in a bad tradeoff. It's obvious with excessive use but I do wonder if it applies even in very minor cass. I look forward to reading your article which seeks to critique locally hosted LLMs specifically, if you do write it.

Feb 21 at 12:08pmWeb
Show threadR3yScaleFeb 21

@glyph
Still, I think that despite the existence of an emerging (not the ones limited to the big tech use of LLMs) argument comphrensively against use of LLM tech, it's still based on what we know about the tech right now. Are we not ultimately limited by our lack of experience? We surely haven't exhausted the possibilities of experimentation with this still new technology, and furthermore our ability to experiment thus far is has been limited by the properatary status of the models and weights, as well as the hardware requirements inherent to the training process.

I get that a lot of people are uninterested in such experimentation or see it as fruitless, and they may end up being right. But why shouldn't Cory or others be able to experiment and experience different possibilities for their workflow, to see how it affects them?

Show threadR3yScaleFeb 21

@glyph
The idea of spellchecking and grammatical standards themselves could be criticised, the way LLM use affects our cognitive abilities could be criticised. But I'm not yet seeing a critque that totally eclipses the possibility that such experimentation might yield an unexpected result?

(of course I'm strictly assuming such experimentation can be done with reasonable sustainability and doesn't involve the current ethical issues in training new LLMs, and I think Cory's current usecase *probably* fits within those parameters)

Show threadR3yScaleFeb 21

@glyph
I wasn't sure about this thread at first but I think it actually has a fair point:
(https://hci.social/@fasterandworse/116104437434039067)

I do think Corys imperative to 'seize' the tech is dubious. Firstly because without open weights we're hardly there, I mean, any experimentation is very limited and the models we would use are permanently stuck in place (which poses concerns about wether embedding open models but closed weights tools in our workflows would induce a kind of material dependency on unethical LLM production after all). But more pressingly, it does imply a that LLMs are somehow inherently valuable and that controlling them would indeed allow us to leap ahead and replace the properietary models. What if there's nothhing to replace and we're better off rejecting the technology?

Still, I just think there might be edge cases, or just unexpected useful areas of application big or small. Like spellcheck and transcription, it has many issues but maybe it's the best we've got until something else comes along, maybe not.

Stephen Farrugia (@[email protected])

Cory Doctorow's ultimate crooked point is that you're fucked if you don't embrace AI. Local models, whatever. He compares technologies that have been proven valuable with a product that is being predicted to be valuable. It's a variation of the assumption that everyone who hates AI hasn't *tried* it. If only they would give it a chance, they'd not be left behind.

🌱 hci.social