Hacker News Reader

Comment by TekMol | Hacker News Reader

Big media companies go fuck yourself.

    That nightmare scenario, for Levin,
    would turn a Food & Wine review into
    a simple text recommendation of a
    bottle of Malbec.

If there is just one wine review on the web which recommends Malbec, AI will not start recommending it. If there are many such reviews, then yes, AI will tell you "Many reviews recommend Malbec.". Just like a human can tell you about what they learned, our artificial brothers have the same right to learn.

Nobody should be allowed to bend the law and extort money from society for writing wine reviews.

View on HN

AI is not my “brother” nor is it a human. This is a predictive model controlled by a for profit business and yea, they should be paying out if they are training on data owned by someone.

It is not a human but it is a fellow intelligent system.

Mankind had this discussion before. When Darwin published his theory about evolution. He faced a lot of hatred because it made humans less special.

Now we go through the same dance again. This time, intelligence in silicon makes humans less special.

It is not intelligent, any more than découpé [1] means scissors and paste are intelligent.

I understand that people like to anthropomorphize things. That's harmless fun when people are imagining tree spirits or thunder gods or whatever. But just because fancy autocomplete produces semi-intelligible text does not mean there is an intelligence behind it. I believe we will eventually be able to create synthetic minds. But our understanding of actual minds is so poor that it's going to be quite a while before we manage.

More than 200 years ago, Mary Shelley told a story of a scientist creating life via the then-new technology of electricity. Since the we keep re-telling the story with the latest technology. It's a good story, and like all good art helps us understand what it means to be human. Which is why I think it's especially bad to use those stories to muddy the many differences between actual humans and something that produces statistically plausible generated text.

[1] https://en.wikipedia.org/wiki/Cut-up_technique

What makes your reply different from a reply ChatGPT would have written?

If nothing - where is that intelligence you are talking about? That intelligence that only humans have?

If you can't tell the difference, I think that says more about you than it does about the text. And I'm definitely not going to try to bring you up to date on centuries of the humanities, or even the decades of cognitive science research. If you'd like to know, perhaps start with a book or two? "The Mind's I", "Kinds of Minds", and "Architects of Intelligence" are accessible places to start.

>they should be paying out if they are training on data owned by someone

Virtually everything you "know" is because of current and past humans. Have you been paying everyone for all of that?

If you want to be a great author and you read books by great writers.. then become a successful writer, are you paying all those authors for training you?

I realize there is a difference with being able to absorb bulk data and all that- however, at the core the argument is still the same.

uh yes? Did you pay for textbooks in school?

Do you regularly copy protected software and claim it as your own creation?

I’ve seen this argument every time someone states that these for profit ai companies should not have complete access to a all of human knowledge and I just don’t see how we’re living in the same world.

It is not the same.

One way to tell is that we think protecting human creators is so important that the US Constitution says Congress has the power to "To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries".

So yes, we absolutely try to pay all those authors. And we also have a variety of cultural mechanisms to reward them in other ways, from honorary doctorates to literary prizes to genius grants to homages. Not to mention a widespread hatred of plagiarism that goes well beyond what copyright covers.

None of that is respected by the current crop of text generators. They chose to release these things knowing of these lacks. And that lack of respect is being richly returned, as it should be.

But you do pay for all of that in todays society.

Teachers are paid, tuition to schools is paid, feedback from tutors is paid, textbooks are paid for, copies of books are paid for, movies and TV are paid for. The experience you learn on the job is paid for via yours or your coworker’s salary.

I get the argument that it’s not copying it’s learning, but it’s also very different than a human learning by observation.

And if learning weights is okay, is it okay to train an AI with the structure designed to reproduce Shrek, then sell the AI? I think ShrekAI would be illegal in most countries. It’s a blurry question that we need to think about.

The training is not the issue. It’s the reproduction. I expect long term the models will still be trained on copyrighted material, but it will refuse to recite any of it.

You don't own data. You can copyright data, maybe.

https://www.americanbar.org/groups/science_technology/public...

That is not correct. LLMs have no inventory of their input datasets. If you give it a billion reviews recommending Malbec, it does not know that many reviews recommend Malbec. It simply has a very good comprehension of what a Malbec review looks like.

In order to know that many reviews recommend Malbec, it needs to be taught this fact specifically, or infer it from context.

Prompt:

Do more Wikipedia pages mention the word "blue" or do more Wikipedia pages mention the word "Accismus"?

ChatGPT:

To determine whether more Wikipedia pages mention the word "blue" or the word "Accismus," we can use a simple search on the Wikipedia website.

Searching for the term "blue" on Wikipedia yields a large number of results, as the color blue is a common topic across various fields. On the other hand, "Accismus" is a relatively obscure term, and it is less likely to be mentioned in a significant number of Wikipedia pages.

Therefore, it is highly probable that more Wikipedia pages mention the word "blue" compared to the word "Accismus."

That is not relevant to what I said. The model can have knowledge about its corpus or tooling to ask questions about it.

But the training process is not a library it can draw upon or answer questions about intrinsically.

>Nobody should be allowed to ... extort money from society for writing wine reviews.

So how do you propose to pay for someone to write wine reviews?

Reply on news.ycombinator.com