OpenAI's Lies and Half-Truths
garymarcus.substack.com
garymarcus.substack.com
Mira Murati knows that this is a controversial viewpoint which is why she's dodging the question. To stretch that out into a whole substack is just the sort of thing Gary Marcus does.
Everyone knows that openAI (and other big AI training company) train on scraped data and copyright works. They claim fair use. Whether or not what they do/have done is fair use is the pivotal question in a whole raft of lawsuits. There's nothing new here as far as I can tell.
[1] eg in the NYT case https://copyrightblog.kluweriplaw.com/2024/02/29/is-generati...
This emphasis on whose blog a bunch of citations is on makes no sense to me, everyone’s got a slant, a discerning customer in the information market is concerned with the signal-to-noise ratio, that was settled at the Labs in 47.
But perhaps more importantly, it mystifies me how a community of people whose lives’ work is substantially if not overwhelmingly IP, and who have been debating “everything should be open” all the way through to “paywalls on blogs should be the norm” and everything in between for almost 20 years now on this forum alone has a big lobby for one of the few trivially bad points in that solution space: intellectual property should be sacrosanct to the point of a net increase in already kinda ridiculous Micky-Mouse-keeps-moving-copyright-goalposts status quo but only for giant megacap corporations and everyone else’s is fair game at arbitrary, automated scale and then a very nifty compression mechanism immediately puts that back into a vault with armed guards.
Open vs closed, patents or no patents or less patents, software licensing, all of this is interesting, all the problems are hard and remain contentious.
It’s impressive in an evil genius way to have somehow figured out a position on this that fails every criteria of every plausible compromised ever proposed to date.
It’s alarming that a forum and community with the mandate and heritage of this one has a sizable lobby that have all pivoted to this whacky, icky place in the search space with the abruptness of an H2O dipole in a microwave.
(put another way, if OpenAI is made to pay for their training data, it is not the small artists that will be paid, it is the megacorps. The small artists will still have the problem of competing with AI, and now those who are competing will be paying a tax to the megacorps to do so, instead of having the option of using their own models. Training as fair use isn't particularly good for smaller artists but the alternative is less good. If you want to improve their lot I think other means are necessary)
My preferred set of legal decisions in the current copyright framework would be: training as fair use, model weights as facts and therefore not copyrightable (though this still annoyingly means models can be kept private), and model output is not copyrightable. This still means artists getting the short end of the stick, but it also severely limits the extent to which a large company can profiteer without a competitor undercutting them (preferably an open-source one, though ATM training foundational models is still sufficiently expensive that's going to be a challenge past the hyper period).
But stronger remedies are possible without many if any new laws: straightforward anti-trust action should have either created housebroken regulated monopolies (e.g. Bell) or broken these companies up years ago. Microsoft alone has fought off such action against a different extremely damning body of evidence and precedent at least once at “front page” levels of plausibility.
The exact point at which a search engine, an OS, a hyper-scaler, a de-facto defense contractor, a hardware OEM, an AdTech platform, … becomes too much scope for anti-market activity is debatable, but it should be pretty uncontroversial that we’re overloading the donkey a bit here with most of the top 10 by cap.
What about, you can’t be a search engine with a copy of the Internet and a secretive mega AI lab? That seems a pretty plausible place to start carving the turkey?
Or, you can’t turn a tax-advantaged 501c-style structure into a de facto subsidiary of such?
Or we create a new category of intellectual property addressing the realities of the modern world in which copyright is too diffuse to let any “derivative work” be anything other than public domain?
The problem would be funding it of course. If only a bunch of hyper-resourced people and institutions felt it was important enough to build AI to create some sort of institution devoted to the public welfare to put the money up for the greater good…
The opposite of the type of person we should be supporting in the tech community.
But it is also true that they are terrible far more than twice a day
It appears to have made some people billionaires. Examples:
Spotify was started by an employee uploading their MP3 collection.
The FB scraped the Harvard student directory.
On the other hand, AI models directly derive their value from the training data, and to a large extent, produces works that are a substitute for the original works.
By the time Google came on the scene, web crawling was well established. AltaVista and Lycos, among others, were already entrenched search engines.
robots.txt was created by the guy that created the first search engine.
Selling human organs will never be the same as selling beef liver, while the difference between a human's liver and a cow's is much smaller than the difference between your brain and a A100 tensor core.
Gen-AI is an empowering technology for the masses and easy to run locally. And the providers are in a race to the bottom on pricing.
(Whether they should be able to hike drug prices by 100% overnight or sell drugs with insane markups is a separate question.)
OpenAI is charging a _really_ high monthly fee for ChatGPT, and it’s quite popular. It’s very limited - there’s no way the cost incurred for usage is near that. Obviously their costs include the R&D that has already happened, but I still think it’s priced way over that.
Drug manufacturers are famously making money hand over fist. Patients aren’t their customers. Insurance companies are. They are surely gouging insurance companies to their fullest capacity.
And big LLM providers will certainly try, but they have competition now, all 3 tied up at roughly the same level. And then there is competition from open models that can handle 50% of what big models do.
So they can only set a high price for very advanced/critical tasks, where usage will be much lower.
With every new technology, there needs to be legislation and judicial rulings to determine the exact definition of things so that the best people can offer you is speculation or their interpretation. The quality of those interpretations will vary by person; I suspect most will be relatively uninformed.
One important factor to consider is the terms of use of the websites from which the data is obtained. For example, YouTube's terms clearly state that content may only be used for personal, non-commercial purposes.
Additionally, the current generation of models does not derive core aspects of knowledge or build world models; instead, they compress vast amounts of data and use that information to generate content. This process likely involves storing copyrighted information, which may violate copyright law.
People produce videos hoping or expecting to be monetized. People who monetize videos (by advertising or sponsoring) do so hoping for the audience to buy some product or service. People who host videos do so expecting to be paid by the monetizing people. That's how the current economic model for ad-supported video works.
A machine learning model is not an economic actor[1]. It's not going to go out and buy a flat wallet or a new set of headphones or whatever else is being advertised on stream. So the current economic model under which that content produced (which is predicated on the assumption that the audience is human) totally falls apart if the audience is an ML model.
[1] yet. And even when they routinely can take part in economic transactions they aren't the audience the advertisers are paying to reach.
The only question left unanswered is how useful OpenAI will be. If it becomes as useful as Google (many claim they already did), then the reality (read: legislation) will bend for them.
After all, while companies didn’t have to pay to pollute the environment, humanity has to eventually suffer; and in the same way the current crop of generative has people’s their work assimilated into models that results in profits for AI companies while the author of the content is denied their fraction of the economic value generated from this process, as well as any recognition and other second order effects they may have benefited from.
So I’ll do it for you, because it’s squarely within site guidelines and norms to defend high-signal comments from getting gang-tackled.
The parent is right, you all know it, and as pernicious and ugly as this side of our business has always been, the stakes are some huge discontinuous amount higher this time.
So downvoters, remember what that button means: it means this is a low-signal or otherwise net-negative value comment on the thread. It explicitly doesn’t mean I disagree.
There’s a disagree button, it’s labeled “reply”. Use it, or get lost.
Regarding the argument, I have seen both sides argued well. The best argument that I have seen for OpenAI being at fault is in cases where gen AI outputs exact copies of copyrighted work, be they visual or text. Dall-E and GPT both need to have serious guardrails added to prevent this, as is shown by the leaked system prompts. If you can jailbrean an LLM, it happily provides the user with exact copies of copyrighted work, in some cases.
The ethical debate had settled into a soft consensus that this sort of thing was icky at best on topics ranging from FB leveraging the Harvard facebooks to Uber ignoring medallion laws to Google scanning all those books to AMP cutting the publisher out of the loop to Amazon eating the OSHA fines rather than give the workers bathroom breaks to FB/Google paying the fines and laughing over EU regulations (which in fairness they seemed to have backed down on), to currently Apple trying to malicious compliance the DMA gatekeeper regulations.
Some of this stuff has been litigated, some not, all of it is cringe-worthy, this is a new low.
The parent cited germane examples from the list. That adds to the discussion.
Folks are free to do the moderate effort thing and write a dissenting reply, they’re not supposed to do the lazy thing and grey out something they spinal-reflex don’t like.
You took the time to write a reply, so even though I think you’re wrong, I applaud you taking the time to state a position.
That's simply incorrect.
And besides, even if garage and dorm room projects sometime do start by illicit means, I don't really think it's responsible to encourage anyone to normaliE that and think it's grossly irresponsible to use that idea to excuse well-connected multi-million dollar organizations who can't hide behind ignorance and should be held to the highest standards lest the playing field for everyone else is ruined.
On your substantial point about garage bands needing to sometimes move fast with limited legal infrastructure but arbitrarily resourced actors being held to a higher standard we’re in violent agreement.
The strongest argument (and it’s not one I’d like to have to defend) for law-breaking tech corporations netting out useful is probably Uber?
The situation pre-Uber was a capture cartel nightmare where basic transportation was a non-starter in certain situations/municipalities.
The (admittedly weak) case that they should still be involved in some, “civil disobedience”, is that the medallion people seem to have decided they’re just never going to give up.
If anyone has a “less bad” example of a big powerful company serving the public good I’d be keen to be reminded of it, but this “fair use for me, copyright for thee” seems to go hand-in-glove with stuff that belongs in both civil and criminal court.
Unfortunately, this is your personal interpretation of the downvote buttons. I agree with your interpretation, in fact, but it is not the norm. And confidently telling people to "get lost" while being misguided is very much not appreciated here.
Here is some comment from the horse's mouth more than a decade ago on this very same topic:
I meant that in the “karma bombing” sense as far as “explicitly” goes.
And I should be more clear in both this subthread (as well as one other comment) that the guidelines I’m referring to aren’t the succinct ones on the literal guidelines page, but rather the evolved apparent consensus under “the @dang administration”.
This is a pretty natural and healthy seeming extrapolation of the idea that rigor should go up as contentiousness goes up, something I personally fail in too often even these days, but will always apologize for failing in.
Threads about OpenAI have set a new record for karma bombing proportional to front-page exposure, blowing away classics like crypto and social media and patent law, I’ve got at least one or two tuples of adjacent comments where the high-quality one got bombed and was on a #1 submission and the far worse form of the argument got too many upvotes on a less-engaged submission.
I’ll decline (for now at least) to speculate on the mechanism/explanation around this (apparent) trend, but it’s inching into “concerning”.
Another take: is real true value creation only possible outside of capitalism? Hence those communist pockets like academia and practices like treating everything as public property are allowed to exist, otherwise there would be no innovation etc just a bunch of companies suing each other?
It’s a slightly more interesting exercise to estimate the multiple on how much OpenAI can’t afford what they’re doing with any conceivable investor if they had to pay for the training data, but I’ll posit it can be demonstrated to the satisfaction of a reasonable person if not to a business-friendly court. I’ll bet the rent it can be demonstrated to a European court.
The semantics of the downvote button are pretty clear on this point.