Google Announces Stealing Part of a Production Language Model | Hacker News Reader

Google Announces Stealing Part of a Production Language Model

twitter.com

117 pointsludovicianul18 comments

I like how they use stealing .... while maintaining steadfast that it's fair use when they use other people's data to train the model.

Ironically, weights and biases are likely not copyrightable.[0] I looked into this first when the LLaMA model was released on BitTorrent, and concluded that downloading it was perfectly legal. The courts even have controlling precedents that say "sweat of the brow" is not a valid justification for copyright, meaning that the effort these large organizations put into training their models does not grant them protection.[1]

[0]https://news.ycombinator.com/item?id=35121093

[1]https://definitions.uslegal.com/s/sweat-of-the-brow-doctrine...

> Ironically, weights and biases are likely not copyrightable. > The courts even have controlling precedents that say "sweat of the brow" is not a valid justification for copyright

You might be right for the US, but the EU seems to have different laws:

https://europa.eu/youreurope/business/running-business/intel...

That's fair enough, but Meta, Google, OpenAI, etc. are all American companies, thus US law would likely apply.

Depends on which part of the world you’re in.

Not when they operate in other countries

IANAL, but I would not immediately agree with the first link, because I believe their reasoning applies to any video encoded with loss.

Not sure what their motivation was to use the word "stealing" here - they are obtaining the output layer weights (mapping from hidden dimension to output), but it seems the hidden dimension size is the most interesting thing here. They could have called it probing rather than stealing.

> Not sure what their motivation was to use the word "stealing" here

Maybe, simply, to get peoples' attention?

Which maps perfectly onto how people would talk about similar work if it wasn't AI.

If I go to the library and read a bunch of books and journal articles to understand something, that's certainly research or education and not "stealing".

If I then use my understanding gained from that education or training to create a new invention, it's plausible that someone could "steal" information from my invention that I never intended to publish, and which was reasonably expected to be private trade secret information not available to others.

The moment that you can train an AI to be a good developer just with a java manual and some slides on design patterns, your analogy will make sense.

But for now an AI requires billion of times more data than a human brain to extrapolate anything useful. Which means that your comparison is meaningless.

I think it's a meaningful comparison still. As a human I can comb through tons of prior art to learn my trade, and that's how we learn most of what we know.

Yes, an AI requires more, but it's still the same basic idea. I'm still not comfortable with using AI to replace artistic functions in particular for a variety of reasons, but the person you're replying to is definitely within the realm of reason.

> it's still the same basic idea

You create weights to decide which word comes next most probably inside your brain and need to do million of multiplications to say a word?

I think it might not be the same basic idea at all.

I think you might want to take a step back and understand the abstraction they were making.

Discussion: https://news.ycombinator.com/item?id=39675735

Better link: https://arxiv.org/abs/2403.06634

I'm so curious what an example set of queries actually looks like here. Like - what were the actual prompts used!

[dupe], discussed previously at https://news.ycombinator.com/item?id=39675735

The twitter post adds no new information and poorly links to the same paper.

Reply on news.ycombinator.com