Comment by imiric | Hacker News Reader

Comment by imiric | Hacker News Reader

Parent Full thread

We must live in alternate realities then.

In the world I live in, Big Tech has unleashed products and services that exploit user data and privacy for profit, enable proliferation of mis/disinformation and propaganda that undermines democracies, enable hostile advertising that manipulates user psychology for more profit, with thinly veiled claims of connecting the world, making the world a better place, and whatever other empty platitude is more marketable at any given moment.

All the while, in general, most services themselves are hardly revolutionizing as their authors claim them to be, and are designed to deliver dopamine hits and keep their users glued to their screens.

We have to stop equating the value of a service to humanity based on its popularity, or the value it generates for its investors. In most cases, the actual disruption technology causes is harmful in ways we have yet to comprehend. We have seen some of it on short-term time scales, but long-term impacts are unknown to even the authors themselves. This is why it's of utmost importance to be conservative with a technology like AI, which unrestrained can send us on the wrong path with no return.

This thread is about the regulation of open source foundation models, not specific uses of them.

If some company uses those tools to create a some kind of product that breaks the law or causes social harm then the law should regulate that specifically. The proposed regulation does a good job of this actually.

But from the way you write, with grand sweeping "we" statements and gesturing vaguely at potential harms and eventual doom, you seem pretty close minded. Perhaps you just don't like that people like these kinds of products and instead of accepting that other people might think differently than you you're self-righteously moralizing about it.

> This thread is about the regulation of open source foundation models, not specific uses of them.

You think opensource models should be immune from the requirement to document how they behave and waht data was used to train them just because they are opensource?

Since training a machine learn model is harmless it should obviously be unregulated open source or not.

If you take it and then do something harmful with it then that should be regulated.

I'm shocked that anyone could think otherwise. Do you have provide documentation on the development process for any other software? Of course not! This is is no different.

> Since training a machine learn model is harmless it should obviously be unregulated open source or not.

For some unspecified definition of harm.

Immediate question: where is the data for learing coming from?

> I'm shocked that anyone could think otherwise. Do you have provide documentation on the development process for any other software? Of course not! This is is no different.

Of course this is different. Since you're training a foundational model that you want others to use everywhere, those who will use it need to know what exactly you trained it with, and where that data is coming from. The users of your data are liable for problems and issues arising from how you trained the model.

E.g. you claim your model is good for using in court proceedings. Somehow you want no responsibility for the outcome of your training even if all you did was train it on screenplays from court dramas. We just have to take your word at face value.

> Of course this is different. Since you're training a foundational model that you want others to use everywhere, those who will use it need to know what exactly you trained it with, and where that data is coming from. The users of your data are liable for problems and issues arising from how you trained the model.

That's exactly my point, if someone wants to use a foundation model they won't use one that they aren't convinced works for their purpose. There is no need for a regulation here. If I open source a model on github but don't document it people won't use it relative to one that is documented.

> E.g. you claim your model is good for using in court proceedings. Somehow you want no responsibility for the outcome of your training even if all you did was train it on screenplays from court dramas. We just have to take your word at face value.

Again, exactly my point. If you are building a lawyer AI bot and use a crappy foundation model as a base the problem is not in the foundation model but in your application. If something bad happens the you the application developer should be held responsible. Which the EU regulation does a good job of regulating and no one is arguing against that part.

The issue with the EU regulation, and your arguments, is that they effect a level too low.

Reply on news.ycombinator.com