Thank you, Sam Altman and everyone at OpenAI, for creating ChatGPT and unleashing the modern era of generative AI, which I use every day to speed up my job and coding at home.
Signed,
Someone who doesn't care that you're making $$$$ from it
Signed,
Someone who doesn't care that you're making $$$$ from it
I understand, Aaron became a martyr; even though he died due to depression and not for "a cause". I applaud what he achieved as a person.
What about if I’m an artist and I don’t want my work included in the training data for an image generation model?
No
You could choose not to publish, and be read
If you are read you can be used to learn from
Copyrights don’t depend on whether I choose to publish a particular work or not.
Zuckerberg personally approved using pirated content to train a model. Is that OK too?
Even Aaron Swartz, the subject of this post, made a distinction (both in his writing/talks and his actions) between academic, scientific, and judicial articles, which he believed should be free for everyone, and other copyrighted content such as books and movies. He seemed to take a softer stance on the latter.
I am not a copyright zealot by any means, but I’m trying to keep an open mind regarding the seemingly knee-jerk “abolish all copyright” takes from certain people in the AI sphere.
Hackernews roots were not about the "original" hacking, but more about Silicon Valley for profit entrepreneurs creating stuff for profit to heavy quotes "make the world a better place" and "democratization X".
The real hackers were in usenet, /. and similar forums. Now, I'm too old and haven't found a good place.
Information wants to be free, copyright is a last century millennium construct to earn money from invention. (Before copyright, people got paid by patrons to invent their stuff).
Copyright should be left at the XX century door, and whoever wants to "monetize" their intellect should use methods for the XXI century.
What I mean is, all laws are fake, but while we have to follow some collective hallucination about magic words giving people authority—to keep society together—the specific delusion that we should enforce artificial scarcity by extending monopolies to patterns of information is uniquely silly.
Walks like a duck and quacks like a duck; I think it's a duck
Once a model is trained, the weights are static unless you train it again. This is a huge practical difference between LLMs and humans. I need to give the LLM the context I want it to consider every single time I use it, because it doesn’t learn from past use at all, it just produces a plausible string of words based on its constant, unchanging set of weights.
At least you're acknowledging that training rights are a proposed expansion of current IP laws!
Not always. That’s more the domain of patents, honestly.
> AI does not replicate your work directly.
This is false, at least in certain cases. And even if it were always true, it doesn’t mean it doesn’t infringe copyright.
> At least you're acknowledging that training rights are a proposed expansion of current IP laws!
Yes, they are, emphasis on the “proposed,” meaning that I believe that training AI on copyrighted material without permission can be a violation of current copyright law. I don’t actually know if that’s how it should be, honestly. But as of right now I think entities like the New York Times have a good legal case against OpenAI.
I think fundamentally we have a difference in opinion on what copyright is supposed to be about. I hold the more classic hacker ideal that things should be free for re-use by default, and copyright, if there is any, should only apply to direct or almost-direct copies up to a very limited time, such as 10 years or so. (Actually, this is already a compromise, since the true ideal is for there to be no copyright laws at all.)
Though... if you say "And the result, despite many years and billions of dollars worth of work, is something that can’t even reliably reference or summarize the material it’s trained on.", doesn't this imply that there's not much to worry about here? I sense that this is a negative jab, but this undermines the original argument that there is so much worth in the model that we need to create new IP laws to handle it.
I mean, I'm not sure what to make of this statement in the first place. Training data should be for the model to learn language and facts, and referencing or summarizing the material directly seems to be out of scope of that. One tends to summarize the prompt, not training data.
No one argued this, to my knowledge. I think that there might be a need for new copyright laws, but the alternative in my mind is that we decide there's not a lot of worth there, meaning that we do nothing, and what OpenAI/Meta/MS/Google/Anthropic/etc are doing is simply de jure illegal. The statement I made about LLMs having major flaws is a point in support of this alternative.
> Training data should be for the model to learn language and facts, and referencing or summarizing the material directly seems to be out of scope of that.
I strongly disagree, as your prompt can (and for a certain type of user, often does) contain explicit or implicit references to training data. For example:
* Explicit: “What is the plot of To Kill a Mockingbird by Harper Lee?”
* Implicit: “How might Albert Einstein write about recent development X in physics research?”
If you take the evolution of those platforms from saying 2005-2015, and project forward ten years, we should be in a much better place than we are. Instead they've gone backwards as a result of enshittification and toxic management.