Post
What is the incremental environmental damage created by running an existing LLM locally on your own laptop?
As to "90% bullshit" - as I wrote, the false positive rate for punctuation errors and typos from Ollama/Llama2 is about 50%, which is substantially better than, say, Google Docs' grammar checker.
enshittification of pluralistic
@tante Since I assume all the #Epstein documents have been scraped into all the LLM models by now, I'd love to see an example of LLM tech being used for good.
Show me the list of Epstein co-conspirators.
Show me names of who helped them escape accountability, and how they did it.
Show me who raped children. Their names, addresses, passport photos.
Then I will believe LLMs and "AI" have delivered a benefit.
Hmmmm... How about this perspective?
LLM is just a programming technique. The ethicality of using LLMs relates to the type of use and the source of the data it was trained on.
Using LLMs to search the universe for dark matter using survey telescopic data or to identify drug efficacy using anonymized public health records is simply using the latest technology for good purpose. Cory's use seems like this.
LLMs trained on stolen data creating derivative work. That's just theft.
Hmmmm... How about this perspective?
LLM is just a programming technique. The ethicality of using LLMs relates to the type of use and the source of the data it was trained on.
Using LLMs to search the universe for dark matter using survey telescopic data or to identify drug efficacy using anonymized public health records is simply using the latest technology for good purpose. Cory's use seems like this.
LLMs trained on stolen data creating derivative work. That's just theft.
@mastodonmigration tagging @pluralistic because this is a good line of discussion and he might need the breath of fresh air you're bringing.
My own two cents: you're missing one of the big complaints in the form of "how they were trained" which is the environment impact angle. Not that it isn't addressed by Cory's use case, just a missing point in the conversation that's helpful to include.
The "stolen data" rabbit hole is sadly a neverending one that digs into deep issues that predate LLMs. Like the ethics of copyright (which is an actual discussion, just so old that it's forgotten in a time when copyright is taken for granted). Using it to create "art" and especially using it to replace artist jobs is however a much much more clear argument.
Nitpick: LLMs can't be used for checking drug efficacy or surveying telescopic data, I think in this line you're confusing LLM with the technology it's based on which is Machine Learning.
Thanks for these corrections. Completely agree with everything, and thanks for tagging Cory.
One of the really unfortunate things that the Silicon Valley scammers have achieved is to coopt new technologies for their despicable pump and dump schemes and apply their disingenuous hype factory which ends up tarring all uses with the same brush.
@mastodonmigration @shiri @pluralistic @tante The only ethical use of a LLM would be one where the training dataset was ethically acquired, the power was minimized to the level of other methods of providing the same benefits, and the 'benefits' were actually measureable and accurate.
None of those are true today, and so far as I know there is little to no path to them.
@reflex @shiri @pluralistic @tante
Seems like Cory's local punctuation and grammer checker is such an example, no?
I completely agree with your view on us being messy, imperfect beings. And while many take such a realization as a free ticket to shrug themselves into deep cynicism, I deeply appreciate people who tend to try a little harder than most to do the right thing, and own every compromise they decide to make as what it is.
Once we start warping our analysis and critical thinking to match our actions instead of trying our best to make our actions fit the former, we'll quickly start losing any ability to be act with any accountability.
@tante I am pursuing what I am calling my "AI Tea Party". It's origin was in putting family birth certificates in a self-hosted tool and realizing I didn't know if that meant they would end up in a training set somewhere. That began a process of me purging direct links to any LLM. Next step is switching to tools that do not use any LLM in their operation. After that is switching from suppliers of anything that use LLMs to operate.
This is a Sisyphean task since these round to **fucking everything** but I'm motivated to pursue it anyway.
@tante We have available listings in your location. You can get our contact number, email address, pictures of house and cost details in this . Please check our listings For more information 🏕️👇
https://site-5272jetqa.godaddysites.com/
@tante Well, I mean, he's wrong, so there's that.
@tante Frankly, I don't think there are any ethical concerns with how he's using it.
The reason AI is a violation when it trains on openly available data and then outputs similar stuff is that it's creating derivative works. Something that reads everything produced by man and then uses that information to score similar output does NOT. It's completely fair use and it's a GOOD application of AI.
IFF all the evil crap that the people who made it are up to wasn't a concern there'd be none.
@tante It seems to me Doctorow is obviously correct about this. But I don't think it matters too much if you don't agree... the trajectory of LLMs is going to be whatever it is going to be.
If you don't like it and have buddies that don't like it either, that's not a bad thing especially if you are undergoing real negative effects from it.
It's just if you stray from reality (whatever that will be) too far for too long, you will end up with a big shock when forced to rejoin it.
Good critique. And I say this, as well as like your arguments, while also believing there are certain domains where LLMs do have reasonable utility value - nowhere near the value required to make OpenAI, Antrophic or the rest profitable, but some value nonetheless.
@tante
I really like and admire @pluralistic and have utmost respect for him, and for precisely those reasons am totally baffled as to why he is claiming "fruit of the poisoned tree" arguments being the cause of LLM scepticism.
The objections to LLMs are to do with what they they are doing right now: destroying the planet, stealing labour, giving power over knowledge to their owners etc.
The objections are nothing to do with LLMs' origins, they're entirely about LLMs' effects in the here and now.
@FediThing @tante @pluralistic Some people - in fact quite a lot; if my reading is correct - do indeed argue that LLMs can *never* be ethically used because they are “trained on stolen work”.
@ianbetteridge @FediThing @tante
Performing mathematical analysis on large corpora of published work is not "stealing."
@pluralistic @ianbetteridge @FediThing @tante If that “mathematical analysis” regurgitates near verbatim works created by other people, it certainly is committing IP theft, and LLMs will happily do that. The “mathematical analysis” is effectively a form of lossy compression on its training data which a prompt can later extract.
@bjn @ianbetteridge @FediThing @tante
Once again, you're talking about *using* a model, not training a model.
Also "IP theft" isn't a thing. Perhaps you mean copyright infringement?
@pluralistic @ianbetteridge @FediThing @tante I’ll give you pedant points for copyright infringement, which is what most people mean by “IP theft”. As for training/using, the difference is somewhat moot. The models are trained to be used, and if trained on copyrighted data without a license, you’ve encoded that data into the model which might then regurgitate it thus facilitating copyright infringement.
@bjn @ianbetteridge @FediThing @tante it is a bedrock of copyright law that devices 'capable of sustaining a substantial non-infringing use' are lawful. Decided in 1984 (SCOTUS/Betamax) and repeatedly upheld.
It is categorically untrue that merely because a model's output can infringe copyright that the model is therefore illegal.
There's not much that's truly settled in American limitations and exceptions, but this is.
@pluralistic @bjn @ianbetteridge @FediThing @tante
I’d argue that It’s a bit more nuanced. Training and inference are two separate stages with their own rules.
For non profit, academic research, excerpts are allowed to be collected, but not the whole work. You still can’t circumvent DRM either.
Llama might be argued is non profit, but lifting whole works to train on still isn’t allowed.
@pluralistic
> untrue that merely because a model's output can infringe copyright that the model is therefore illegal.
Mhmmm naaah overfitting and memorization are very much a thing, especially in the case of LLM where they've completely given up on controlling data leaks, and where memorization has been demonstrated rather unambiguously e.g. with the suitesparse example...
Not to imply that "illegal" is bad ofc, or that copyright justifiable
@bjn @ianbetteridge @FediThing @tante 'facilitating copyright infringement' just isn't a thing.
@bjn @ianbetteridge @FediThing @tante and as befits UK fair dealing (and related limitations and exceptions), we've had opinions from IPREG affirming that training a model doesn't infringe.
@pluralistic
> IPREG affirming that training a model doesn't infringe.
What we now take the party line serious?
@pluralistic @ianbetteridge @FediThing @tante Then the laws are not fit for purpose. The whole point of copyright is to encourage people produce works by being sure they get the benefit of those works. If my works can be encoded into a bunch of matrix weights and reproduced without attribution let alone financial recompense, then why should I bother? Google is doing its best to effectively steal the bread out of creators mouths with its AI summaries. It may be legal, but it stink.
@bjn @ianbetteridge @FediThing @tante by all means say 'i don't like this technology' but don't conflate that with 'therefore it is illegal'
@pluralistic @ianbetteridge @FediThing @tante Well apart from Anthropic having to pay $1.5B for copyright infringement, it’s all above board, 🙄. It’s not a matter of liking the technology or not, machine learning is capable of cool and useful things. However, how LLMs are being used and pushed is both immoral and culturally destructive. I’m surprised you are buying into it.
@pluralistic @ianbetteridge @FediThing @tante “Mathematical analysis” is doing a lot of work here. It could mean gathering meaningless statistics. Or it could mean capturing the qualities (deviations from the average) that make a particular work of art (or author) special, creative, surprising—for use in simulacra.
I think that's harmful, to the culture as a whole, if not to the artworks and artists getting regurgitated.
@gleick @ianbetteridge @FediThing @tante
Let's stipulate to that (I don't agree, as it happens, but that's OK). It's still not a copyright infringement to enumerate and analyze the elements of a copyrighted work.
For the record, I think AI art is bad and neither consume nor make it.
@pluralistic @gleick @ianbetteridge @FediThing @tante there's been documented cases of LLMs regurgitating stuff from their training set verbatim, which clearly IS copyright infringement; and that means some parts of the training set are.encodrd in the weights of the model, which looks like publishing a copyrighted work to me. If publishing a JPEG of an image without copyright to it would be infringing, isn't publishing a model that can recreate something also infringing?
@pluralistic @ianbetteridge @FediThing @tante I'm not claiming that's copyright infringement. Even if one respects the general framework of copyright, which I know you don’t, it seems hopeless to apply it to this AI mess.
But there is a kind of theft here. Not that it's actionable or measurable. But it’s nontrivial. It's related to questions of impersonation. It's an assault on individuality. Whatever your reasons for thinking AI art is bad (I have some sense), it's related to that, too.
@pluralistic @ianbetteridge @FediThing @tante Some authors have taken the view that they deserve some compensation for the use of their books in training the LLMs. Do the transaction and we're hunky-dory. That's not my view. I don't care about compensation; I just don't want my prose regurgitated in the LLMs, for reasons I'm not yet able to express properly. I feel I should have been asked, and I feel violated.
I agree with you, I feel the same way and I have not been published.
I’m really enjoying The Three Ages of Water. I slowed down my read to really enjoy your writing and the amazing breadth of knowledge in it.
@schmubba Thank you, and I liked it, too, but I didn't write it. That was @petergleick.
@gleick @pluralistic @ianbetteridge @FediThing @tante It's the regurgitation part. People read your work all the time, and are inspired by it to create their own works. I'm sure you're fine with that. But when they read your work and proceed to regurgitate chunks of it and claim it as their own? Because that's what the LLMs are doing all too often, and the reasons to object to LLMs doing it are the same as the ones to people doing it.
@gleick @pluralistic @ianbetteridge @FediThing @tante I think the sense of “theft” that creators feel is directly caused by the fact that the AI industry (as it stands today) is a Ponzi scheme which is fundamentally built on remixing creators’ works. I have a feeling that most creators will not feel the same kind of outrage if an educational institution created the same technology for academic use, to generate insights into online culture and psychology.
In short, the GRIFT (i.e. the particular application of the technology) is the source of the feeling of theft, not the technology itself. I think the tech itself has value when used ethically.
FWIW I agree with Cory here that copyright is the *wrong* framework to use for criticizing AI, because for every caes where copyright helps the individual creator, there are hundreds of times where it helps incumbent megacorporations more.
@gleick @ianbetteridge @FediThing @tante
Let's stipulate to that (I don't agree, as it happens, but that's OK). It's still not a copyright infringement to enumerate and analyze the elements of a copyrighted work.
For the record, I think AI art is bad and neither consume nor make it.
@pluralistic @FediThing @tante I agree - but that is the (incorrect imho) argument they make.
@pluralistic @ianbetteridge @FediThing @tante It really depends a bit on the details, doesn't it. If I copy a CD, I also perform some mathematical analysis on it, error checking etc. Maybe I even make non-exact copy by passing it through some filter to make it sound better. But it's totally different from listening to a bunch of Beatles songs and then getting inspired to write my own songs in a similar style.
@pluralistic @ianbetteridge @FediThing @tante
It's still profit loss damage curable by income transfer if the illegally acquired data was used to create that profit. Dataset prominence should provide the percentage of profits and prominence is data size but also inference casualty. The primary literature should not be able to be diluted with free intellectual property.
I don't know if any of this is actual case law and I'm not a lawyer.
@drdrowland @ianbetteridge @FediThing @tante
You're talking about ways of using models, not the creation of models. It's possible to make a model that does illegal things. But training a model is not illegal.
@FediThing @tante @pluralistic I’ve been utterly baffled why he’s so popular for decades.
@Colman @FediThing @pluralistic this is just a dumb attack on Cory as a person which I will not accept. You can talk about what he or I wrote (both things can be criticized) but have some respect
@Colman @FediThing @tante That's interesting. I've never wondered that about you.
@pluralistic capital murder on the timeline
@pluralistic @Colman @FediThing @tante
This is...disappointing. To be fair, I'm disappointed in almost everyone in this thread for engaging in schoolyard shit throwing, but you're much higher in status and your shit sticks. Have a conversation. Figure out where these views can comingle. Find common understanding or you risk using your high status to fracture an already unstable alliance of people who want technology to operate safely and for the benefit of our shared humanity.
Do better.
Which parts of running a model on your own laptop are implicated in "destroying the planet?" How is checking punctuation "stealing labor?" Or, for that matter "giving power over knowledge to LLM owners?"
@pluralistic @FediThing @tante The problem with AI is not primarily with tech itself, but human traits of laziness, greed and ignoring unpriced externalities for short-sighted personal gains. Those of course exist even without AI, but AI allows for their damage to be multiplied many thousandfold which upgrades it from minor annoyance to existential crisis.
And the problem with Cory saying he's preferring using his AI is not exact cost of him using AI model (which is tiny), it is NORMALIZING 1/x
@pluralistic You should read the top-of-thread link. It lays out in detail the author's disagreements with you in a respectful manor.
Maybe you'll call it incoherent because you don't want to listen to accusations of going against your own principles, noone does. But I would disagree.
@pluralistic i'd start with the part that the model probably came pre-trained. Or was it trained by you on your laptop...? @FediThing @tante
This is a purity culture argument about the "fruit of the poisoned tree." The silicon in your laptop was invented by a eugenicist. The network your packets transit was invented by war criminals. The satellite the signal travels on was launched on a rocket descended from Nazi designs that were built by death-camp slaves.
@pluralistic i guess this misses the point: the particular chip in my laptop wasn't made by war criminals (i hope...), but the model you do use was trained under vast amounts of energy and water consumption. I'm not sure this is completely comparable, tbh.
No, this is just more "fruit of the poisoned tree" and your argument that your fruit of the poisoned tree doesn't count is the normal special pleading that this argument always decays into.
@pluralistic sorry, i'm just not good at making a point. To me, not "LLM" is the "forbidden fruit", but "using an LLM for certain purposes" is. I think there are actually use-cases for stochastic inference machines (like folding proteins or structuring references), but, as @tante wrote (better: as I understand him), there are use-cases that one very much can reject in its entirety. And that should be okay.
I never denied the existence of "use-cases that...one can reject it its entirety."
@pluralistic and yes, i'm aware that producing a chip also costs vast amounts of energy and water... but at least my chip is used to solve a multitude of purposes, while a LLM that checks spelling and grammar is built and trained for one single use-case (that, nb, could also be done without an LLM). So yes, I do differenciate. @FediThing @tante
Llama 2 was not built to check spelling and grammar. That's "not even wrong."
To be clear, I completely reject this argument as a form of special pleading. Everyone has a reason why *their* fruit of the poisoned tree is OK, but other peoples' fruit of the poisoned tree is immoral.
@pluralistic Think a bit harder about these questions yourself.
Suppose you wore a coat made out of mink fur. The minks are already dead, simply wearing the coat won't kill more minks. What does wearing mink fur have to do with cruelty to minks?
Suppose you live in the time of the Luddites. Legislation prohibits trade unions and collective bargaining. Mill owners introduce machines, reducing wages. But you build your own machine. Problem solved? You helping labor or capital?
This is a "fruit of the poisoned tree" argument.
Suppose you use a computer to post to Mastodon, despite the fact that silicon transistors were invented by the eugenicist William Shockley, who spent his Nobel money offering bribes to women of color to be sterlized?
Suppose you sent that Mastodon post on a packet-switched network, despite the fact that this technology was invented by the war criminals at the RAND corporation?
@pluralistic I don't think mink fur or LLMs are comparable to criticizing the origins of the internet or transistors. It's the process that produced mink fur and LLMs that is destructive, not merely that it's made by bad people.
For example, LLM crawlers regularly take down independent websites like Codeberg, DDoSing, threatening the small web. You may say "but my LLM is frozen in time, it's not part of that scraping now", but it would not remain useful without updates.
No. Literally the same LLM that currently finds punctuation errors will continue to do so. I'm not inventing novel forms of punctuation error that I need an updated LLM to discover.
@pluralistic Ok, fair enough, if spell checking is literally the only thing you use LLMs for.
I still think you wouldn't rely on a 1950s dictionary for checking modern language, and language moves faster on the internet, but I'm willing to concede that point.
I still think a deterministic spell checker could have done the job and not put you in this weird position of defending a technology with wide-reaching negative effects. But I guess your post was for just that purpose.
I'm not using it for spell checking.
Did you read the article that is under discussion?
@pluralistic I apologize, I did in fact read the relevant section of your post, and I was using spell-checking as shorthand for all typo checking, because deterministic grammar checkers have also existed for some time, although not as long as spell checkers and perhaps they have not been as reliable. I understand that LLMs can catch some typos that deterministic solutions may not.
I just think we should put more effort into improving deterministic tools instead of giving up.
@pluralistic@mamot.f I don't think mink fur or LLMs are comparable to criticizing the origins of the internet or transistors. It's the process that produced mink fur and LLMs that is destructive, not merely that it's made by bad people.
For example, LLM crawlers regularly take down independent websites like Codeberg, DDoSing, threatening the small web. You may say "but my LLM is frozen in time, it's not part of that scraping now", but it would not remain useful without updates.
Also, you're wrong about the Luddites, just as a factual matter. The guilds the Luddites sprang from weren't prohibited by law, they were *protected* by law, and the Luddites' cause wasn't about gaining new protections under statute, but rather, enforcing existing statutory protections.
(Also: the Luddites didn't oppose steam looms or stocking frames; their demands were for fair deployment of these)
@pluralistic Thank you for the fact check. I was paraphrasing that text from the popular Nib comic: https://thenib.com/im-a-luddite/
If this contains factual inaccuracies I will need to do more research and perhaps stop sharing that comic.
@skyfaller @FediThing @tante I strongly recommend Brian Merchant's "Blood in the Machine" as the best modern history of the Luddites.
(Hello Mr Doctorow! Just want to make clear I admire you a great deal and this isn't intended as an attack on you!)
Running a local LLM with no connection to outside providers might be a way of avoiding bad stuff, but I am not clear on how this relates to discussing origins of technologies?
It seems like there's ambiguity in your post about whether it applies just to people with homelabs wondering if they should try offline LLMs, or whether you are discussing LLMs as a general technology?
Almost everyone using LLMs will use the online kind, so objections to LLMs are (reasonably IMHO) based on that scenario.
> I am not clear on how this connects to discussing origins of technologies
Because the arguments against running an LLM on your own computer boil down to, "The LLM was made by bad people, or in bad ways."
This is a purity culture standard, a "fruit of the poisoned tree" argument, and while it is often dressed up in objectivity ("I don't use the fruit of the poisoned tree"), it is just special pleading ("the fruits of the poisoned tree that I use don't count, because __").
@pluralistic @FediThing @tante
What's the difference between your argument here and "Slavery is OK because I didn't kidnap the slaves; I just inherited them from my dad." ??
Because there are no slaves in this instance. Because no one is being harmed or asked to do any work, or being deprived of anything, or adversely affected in *any articulable way*.
But yeah, in every other regard, this is exactly that enslaving people.
Sure.
@pluralistic @FediThing @tante
Unless you consider stolen intellectual property (and ongoing copyright violations) a harm, a deprivation, &c.
But your general analogy against "fruit of the poison tree" morality would seem to also apply in the case of slavery -- in my hypothetical, the person didn't enslave anyone. They just inherited a slave from someone who did. That is indeed "fruit of a poisoned tree", even if they just continued an existing enslavement.
We have a real world recent example -- the cell lines stolen from Henrietta Lacks. Do you dismiss any moral concerns about using her cell line without consent as a neo-liberal moral purity trap?
Scraping and training are not copyright infringements:https://www.theguardian.com/us-news/ng-interactive/2026/jan/18/tech-ai-bubble-burst-reverse-centaur
Thank you for the responses 🙏
"Because the arguments against running an LLM on your own computer"
...ahhh okay. So was this post aimed more at a very narrow homelab kind of audience?
It's just, as a reader, the article's emphasis on examples of tech origins imply it's trying to defend LLMs in general? This probably is my ignorance as a reader, but it's how it came across to me, and led to bafflement.
@FediThing @tante This is the use-case that is under discussion.
@pluralistic @FediThing @tante So this is some kind of spell-checker, which is already in LibreOffice? I'm not sure why I would use that instead.
I use other kinds of offline LLMs, especially for visual effects, for subtitles, fixing dialogue errors, auto mixing, etc. There are "deep fake technologies" useful for mocap, camera tracking, and such other useful techs. They don't use prompts, and don't generate art.
Perhaps we need a new name to differentiate it from the online genAI tech.
@pluralistic@mamot.fr @FediThing@chinwag.org @tante@tldr.nettime.org you're attempting to legitimize use of an unethical technology for something you don't actually need a plausible-sounding-wall-of-text generator for
it goes beyond "it's made by bad people in bad ways". it's a ""tool"" that actively causes cognitive decline and psychosis and sucks the soul out of everything it touches. and mind you promoting and legitimizing it is an act of support for those bad people and their bad ways. your deflection is a typical that of someone with no regard for ethics
"I installed Ollama" instantly gives a person away as a techbro
I'm not a liberal, I'm a leftist, so perhaps this is why I disagree with you.
The argument that "something is unethical because someone else used it in an unethical way" is so incoherent that it doesn't even rise to the level of debatability.
Thanks. Can totally see how that makes sense at a technical level for people who run their own offline services.
I think it's the ambiguity that is driving the discourse over this post. People are taking the "refusing to use a technology" section as a defence of LLMs in general?
If the angle was caging LLMs or something like that it, might make it clearer that you aren't endorsing the most common form of LLM?
Anyway, it's your call on this as author, just wanted to feed back on this because your writing matters and I hope feedback is helpful to it.
@FediThing @pluralistic @tante i feel in the similar way as big tech has taken the notion of AI and LLMs as a cue/excuse to mount a global campaign of public manipulation and massive investments into a speculative project and pumps gazillions$ into it and convinces everyone it's innevitable tech to be put in bag of potato chips, the backlash is then that anything that bears the name of AI and LLM is poisonous plague and people are unfollowing anyone who's touched it in any way or talks about it in any other way than "it's fascist tech, i'm putting a filter in my feed!" (while it IS fascist tech because it's in hands of fascists).
in my view the problem seems not what LLMs are (what kind of tech), but how they are used and what they extract from planet when they are used by the big tech in this monstrous harmful way. of course there's a big blurred line and tech can't be separated from the political, but... AI is not intelligent (Big Tech wants you to believe that), and LLMs are not capable of intelligence and learning (Big Tech wants you to believe that).
so i feel like a big chunk of anger and hate should really be directed at techno oligarchs and only partially and much more critically at actual algorithms in play. it's not LLMs that are harming the planet, but rather the extraction, these companies who are absolute evil and are doing whatever the hell they want, unchecked, unregulated.
or as varoufakis said to tim nguyen: "we don't want to get rid of your tech or company (google). we want to socialize your company in order to use it more productively" and, if i may add, safely and beneficialy for everyone not just a few.
@prinlu @FediThing @pluralistic @tante I agree with most things said in this thread, but on a very practical level, I'm curious what training data was used for the model used by @pluralistic 's typo-checking ollama?
for me, that training data is key here. was it consensually allowed for use in training?
because as I understand, LLMs need vast amounts of training data, and I'm just not sure how you would get access to such data consensually. would love to be enlightened about this :)
@bazkie @prinlu @FediThing @tante
I do not accept the premise that scraping for training data is unethical (leaving aside questions of overloading others' servers).
This is how every search engine works. It's how computational linguistics works. It's how the Internet Archive works.
Making transient copies of other peoples' work to perform mathematical analysis on them isn't just acceptable, it's an unalloyed good and should be encouraged:
https://pluralistic.net/2023/09/17/how-to-think-about-scraping/
@pluralistic @tante After reading so many comments, it is pretty clear who here would be opposing the creation of Napster and torrenting and be defending RIAA... They are also clearly very much against Internet Archive, shadow libraries, etc, simply because they can't take any disagreement.
Who knew running a local LLM, that uses the same energy as watching a youtube video, to spellcheck your own work would bring out such a mob.
@pluralistic @prinlu @FediThing @tante I think the difference to search engines is how LLM reproduces the training data..
as a thought experiment; what if I'd scrape all your blogposts, then start a blog that makes Cory Doctorow styled blogposts, which would end up more popular than your OG blog since I throw billions in marketing money at it.
would you find that ethical? would you find it acceptable?
further thought experiment; lets say you lose most of your income as a result and have to stop making blogs and start flipping burgers at mcDonalds.
your blog would stop existing, and so, my copycat blog would, too - or at least, it would stop bringing novel blogposts.
this kind of effect is real and will very much hinder cultural development, if not grind it to a halt.
that is a problem - this is culturally unsustainable.
@bazkie @prinlu @FediThing @tante
First: checking for punctuation errors and other typos *in my own work* in a model running on *my own laptop* has nothing - not one single, solitary thing - in common with your example.
Nothing.
Literally, nothing.
But second: I literally license my work for commercial republication and it is widely republished in commercial outlets without any payment or notice to me.
@pluralistic but then you consented to that, right? you are in control of that.
also my example is very similar - after all, it's data scraped without consent, used to create another work. the typo-checker changes your blogpost based on my training data, in the same way my copycat blog changes 'my' works based on your training data.
Sure, it's on a way different scale - deliberately, to more clearly show the principle - but it's the same thing.
Should we ban the OED?
There is literally no way to study language itself without acquiring vast corpora of existing language, and no one in the history of scholarship has ever obtained permission to construct such a corpus.
Dictionaries reference the sources they use for examples in the entries themselves.
LLMs lose the references at training time.
You've got this dead wrong.
@pluralistic
Does the OED provides references to the quotes?
Does an LLM provides references to the data its weights are trained on?
These things are not the same. You keep repeating the same non-argument. Check your own work with a model running on your laptop. Don't call a bunch of people purists and when they take offense, try to justify that with this nonsense about dictionaries.
@pluralistic I gave it a good thought, and you know what, I'm gonna argue that yes, for me there is a degree of unethical-ness to that lack of permission!
the things that makes me not mind that so much are a variety of differences in method and scale;
(*btw just explaining my personal reasons here, not arguing yours)
- every word in the OED was painstakingly researched by human experts to make the most possible sense of it
- coming from a place of passion on the end of the linguists, no doubt
- the ownership of said data isn't "techno-feudal mega-corporations existing under a fascist regime"
- the OED didn't spell the end of human culture (heh) like LLMs very much might.
so yeah. I guess we do agree that, on some level, the OED and an LLM have something in similar.
it's the differences in method and scale that make me draw the line somewhere in between them; in a different spot from where you may draw it.
and like @zenkat mentioned elsewhere, it's the whole thing around LLMs that makes me very wary of normalizing anything to do with it, and I concede I wouldn't mind your slightly unethical LLM spellchecker as much, if we didn't live in this horrible context. :)
I guess this has become a bit of a reconciliatory toot. agree to disagree on where we draw the line, to each their own, and all that.
@pluralistic @bazkie @prinlu @tante
This would be my take:
Search engines direct people to the work they index. They reward labour by directing people towards it.
Scraping without consent for training data lets people reproduce the work without crediting or rewarding the people who actually did the labour. That seems like labour theft?
If it is labour theft, then it isn't sustainable and that's part of why LLMs are so questionable as a technology.
@FediThing @bazkie @prinlu @tante
There are tons of private search engines, indices, and analysis projects that don't direct text to other works.
I could scrape the web for a compilation of "websites no one should visit, ever." That's not "labor theft."
@pluralistic @bazkie @prinlu @tante
Indexing works is a totally different thing to creating knock-offs of works, surely?
What Miyazaki said about AI knock-offs surely illustrates the difference?
@FediThing @bazkie @prinlu @tante
No one is defending "creating knock offs of works." Why would you raise it here? Who has suggested that this is a good way to use LLMs or a good outcome from scraping?
@pluralistic @bazkie @prinlu @tante
If LLMs were only used for checking grammar that is one thing.
But by far the most common use of LLMs is labour theft through creating knock-offs, and that's something else.
I think the concern is that training data useful for the first case could be useful for the second case too? Hence the questions about where the training data comes from and where it ends up.
Kind of feels like it needs to be strictly ringfenced if it's to be ethical?
@FediThing @bazkie @prinlu @tante
Once again, you a replying to a thread that started when someone wrote that using an LLM to check the punctuation in your own work is ethically impermissible because no one should assemble corpora of other peoples' works for analytical purposes under any circumstances, ever.
@pluralistic @FediThing @prinlu @tante sure, but I'm responding here specifically to your statement that scraping for training isn't unethical per se.
@FediThing @bazkie @prinlu @tante
The argument was literally, "It's not OK to check the punctuation in *your own work* if the punctuation checker was created by examining other peoples' work, because performing mathematical analysis on other peoples' work is *per se* unethical."
@pluralistic
The argument was "without the consent of the creators of said works." And you know that.
Don't be just another debate bro. Please.
@pluralistic @FediThing @prinlu @tante I'd say "because performing [automated, mass scale] mathematical analysis on other peoples' work [without their consent] [with the goal of augmenting one's own work] is *per se* unethical" - and in that case, it's a statement I would agree with.
@bazkie @FediThing @prinlu @tante
You've literally just made the case against:
* Dictionaries
* Encyclopedias
* Bibliographies
And also the entire field of computational linguistics.
If that's your position, fine, we have nothing more to say to one another because I think that's a very, very bad position.
@pluralistic @bazkie @FediThing @prinlu @tante encyclopedias don't say: look i made a painting, wanna buy?
but all that is kinda offtopic, imo
there are plenty of reasons to not use LLM, besides that commons/copyright stuff, wich are not purist and very much based on real-world-issues. in a perfect world, theft won't be an issue (imo), because we had overcome money and fossil energy. but using genLLM still would be morally wrong, because of its dangers, due to bias and failures. …/
@pluralistic
No, because dictionnaries are about language which is a shared common, encyclopedias are about knowledge, which is a shared common, and bibliohraphies are a list of works, not a derivative.
Knowledge, language and a list of works cannot be copyrighted. You can use language, knowledge, words from the dictionary. You can quote an encyclopedia when refering to the source. None of that is even relevant to this discussion.
@pluralistic @bazkie @FediThing @prinlu @tante I think part of the issue here is that GenAI is being pushed so hard and fast *everywhere* that's it's hard to be nuanced about what narrow use-cases might be acceptable or not.
We're living under a massive pro-LLM propaganda campaign. They have already set the terms of the debate with a maximalist position. It's no surprise that the backlash is similarly absolute.
@pluralistic @FediThing @prinlu @tante I did not make that case, if you'd properly read my [additions] to that statement. making dictionaries etc isn't automated on mass scales like feeding training data to LLMs is.
it's a very human job that involves a lot of expertise and takes a lot of time.
@FediThing @bazkie @prinlu @tante
By this standard the OED is unethical.
> Almost everyone using LLMs will use the online kind, so objections to LLMs are (reasonably IMHO) based on that scenario.
Except that in this specific instance, you are weighing on an article that claims that it is wrong to run a local LLM for the purposes of checking for punctuation errors.
Install bonfire.cafe
Get the full app experience
