Instead of defending the use of LLMs for polishing up your writing, we could be advocating for unpolished writing. Blog posts with spelling errors and awkwardly repeated words. Emails that sound a bit less warm and professional because you forgot the preamble of "Apologies for the late reply, hope you're well! Thanks for the thing last week".

If there's no budget for a human editor, why should the text meet a "professional" (middle class, formally educated) standard? Dyslexic people can just write how they write and people can deal with it. Autistic people can just say what they mean to say and not waste energy on the double empathy gap.

We can learn to read for a more inclusive world, instead of wasting the planet's diminishing resources masking our differences.

@zoy 'instead of accommodations we should just eliminate the role of grammar in class distinction.' this is my least favourite leftism (and I'm a socialist). no reform, only radical change allowed! people asking for incremental accommodations are traitors to the revolution!

making writing accessible to all is one of the very few good things about LLMs. let the unconfident writers polish their prose and let's target the financial and environmental and industrial injustice and unsustainability of 'AI' directly.

@onekind @zoy writing was already accessible to all if you don’t apply unreasonable and unnecessary standards.

LLMs are not accessible to all. They incur dependency on one of a handful of massive mostly-US corporations. And prices are almost certainly going to start increasing because current prices are artificially low and insufficient to cover operating costs (presumably in order to inflate adoption rates).

LLMs aren’t a socialist solution in the slightest.

@benjamineskola @zoy wow that's some interesting tense work. 'was already' sounds like 'is' right up until you jam in the 'if' that begs the question i was posing.

writing is open to anyone but writing *accessibly to others* is difficult. I love the idea that we should all tolerate unconventional writing right up until I remember that *readers also need to be able to understand it.*

I am pretty fucking sure my toot made it clear I'm not pro-LLM and even laid out the bases for a socialist critique of them so maybe stop putting words in my mouth.

@onekind You did, indeed, say "it's one of the very few good things about LLMs", not that LLMs as a whole are good. But I disagree: it's not a good thing about LLMs. It's the opposite of a solution to the problem; it makes the problem worse.

Even if LLMs aided communication, they do so at the cost of increased dependency on US tech corporations (or, at best, non-US tech corporations).

And I don't agree even that they do aid communication. Inserting a text generator in between a writer and a reader is the opposite of communication.

(And I didn't put words in your mouth.)

@benjamineskola You wrote 'LLMs aren’t a socialist solution in the slightest,' negating something I never claimed, and that's putting words in my mouth.

Yet again an 'if' that concedes my point only to argue things *should* be otherwise. Yes they should, but they are not anywhere near close to being that way.

You're missing something important in your critique of LLMs: there are thousands of them, some are open source, many will run on your own hardware, and many can be prompted regarding tone and style. There is nothing inherently American or even capitalist about GPT or LLMs per se. The tech can be used in prosocial and pro-socialist ways, on multiple sides of struggle.

@onekind OK, a few things: first, no, that's not putting words in your mouth. I am asserting that, if someone thought LLMs are compatible with socialism, they'd be wrong. (As it stands in the present moment.)

As for the 'things should be otherwise', well, yes. We are socialists; the whole point is that we want things to be otherwise. I don't think it makes sense to draw some lines and insist that this particular thing can't be changed, when we're already demanding fundamental changes in economic and social arrangements.

As for the point about 'open-source' LLMs: that doesn't actually avoid the problem. They aren't actually open-source, and training the weights requires the same intensive process as is done for the closed-source LLMs. Nobody can do that without millions of dollars to burn, and the existing ones all depend on the training done by those same corporations. Better/fuller explanation: https://hachyderm.io/@david_chisnall@infosec.exchange/116107854327461213

I do agree that in a socialist society the technology itself could potentially exist and be used without contradiction. But in the real world, in the present, the only possible usage of LLMs entail that dependency on capital.

Hachyderm.io

@benjamineskola There is nobody else in this conversation; negating a claim imputes it to me.

And it's pointless to assert an ought without any sense of the weight massed against it or any realistic program for change.

Which you tacitly acknowledge because you're making exactly the same argument back to me — that there's no point asserting any prosocial or socialist angle to LLMs because of path dependency and their deeply embedded capitalist origins. Audre Lord, the master's tools, etc.

There is absolutely nothing to stop a government somewhere deciding to train up a model on a more multicultural corpus. The technology is not inherently one thing or another. The implementations we've seen to date have come out of a particular corporate conjuncture but that governs their present not the future. As noted by Foucault — the great scholar of the otherwise — power is always exercised in struggle and remains constantly in danger of reversal.

And the monetary expense of doing so is absolutely nothing compared to the difficulty of changing every aspect of social existence in which literacy implies class credibility and cultural capital. I do behaviour change for a living and it is a slow grind getting people to adopt even things that will benefit them.

Anyway it's 3AM here so I'm going to leave you to it. Thanks for the chat.