Hacker News Reader

Comment by andy99 | Hacker News Reader

Most importantly, 2023 was the year when "open source" got watered down to mean "you can look at the source code / weights" if you agree with a bunch of stuff. Most of the models referenced, like stable diffusion (RAIL license) and Llama & derivatives (proprietary facebook license with conditions on uses and some commercial terms) are not open source in the sense that it was understood a year ago. People protested a bit when the terminology started being abused, and now that's mostly died down and people now call these restrictive licenses open source. This (plus ongoing regulatory capture) is going to be the wedge that destroys software freedom and brings us back to a regime where a few companies dictate how computers can be used.

View on HN

Mistral 7B [1] and many models stemming from it are released under permissive Apache license.

Some might argue that a "pure" open-source would require the dataset and the training "recipe" as it would be needed to reproduce the training, but it would be so expensive that most people wouldn't be able to do much with it.

IMO, a release with open weights without the "source" is much better than the opposite, a release with open source and no trained weights.

And it's not like there was no progress on the open dataset front: - Together just released RedPajama V2, with enough tokens to train a very sizeable base model. - Tsinghua released UltraFeedback which allowed more people to align models using RLHF methods (like the Zephyr models from Hugging Face) - and many many others

[1] https://mistral.ai/news/announcing-mistral-7b/ [2] https://github.com/togethercomputer/RedPajama-Data

> IMO, a release with open weights without the "source" is much better than the opposite, a release with open source and no trained weights.

Why not both?

Isn’t everyone using data sets they don’t have rights to? Or at least not right to release.

Both would be even better but it would be much smaller progress for the open source community in comparison to the release of the weights.

Even if they don't release the (copyrighted) training data itself, do they provide a "recipe" to reproduce it?

Like, a list of sources used and how they were harvested?

It really depends on the release. It's not the case for the base Mistral 7B model. But many finetuned models are released with some information about the dataset used.

How about "Open Binary"?

> Some might argue that a "pure" open-source would require the dataset and the training "recipe" as it would be needed to reproduce the training, but it would be so expensive that most people wouldn't be able to do much with it.

I've actually argued the opposite. https://www.marble.onl/posts/considerations_for_copyrighting... When you look at the freedoms underlying open source, you can exercise them without the training data.

Most of what disqualifies the licenses I mentioned above from being classically open source is use restrictions.

Yes, I think most of the misunderstanding comes from people who are not in ML. A lot of the times you can't release the training pipeline without getting into grey areas of training data copyright. A ML model is a high dimensional distribution that can be sampled from infinitely, given enough time and money you can train another model (possibly even with a different architecture) that closely models the original model in generation capability.

In practice this matters less than you think. You can’t easily prove that any outputs were generated by a particular model in general, so any user can simply ignore your licenses and do as they please.

I know it rustles purist feathers, but I don’t understand why we live in this pretend world that assumes that folks particularly care about respecting licenses. Consider how little success that the GNU folks have had with using the courts for any enforcement of their licenses, and that’s by stallmans own admission.

AI is itself a subversive technology, whose current versions rely on subversive training techniques. Why should we expect everyone to suddenly want to follow the rules when they read a poorly written restrictive open source license?

For personal or noncommercial use I agree the restrictions are meaningless. As they are for "bad actors" that would potentially abuse the tools in contravention of the license. But the license terms are a risk for commercial users especially when dealing with a big company like Meta. These risks werent previously there in say pytorch that is MIT licensed. The ironic thing with these licenses is that they are least enforceable on those who would be most likely to abuse them: https://katedowninglaw.com/2023/07/13/ai-licensing-cant-bala...

Re success of free licenses, linux (other than a few arguable abuses) has remained free and unencumbered thanks to GPL licensing.

Somehow the "AI defense" (namely that it is not possible to "prove" anything was used illegally) will open Pandora's box in terms of providing viable channels for whitewashing explicit theft activity. Steal anything proprietary, run it through an AI filter that mixes it with other stuff and claim it as your own.

A human could have done that at any point in the past. The only novelty is now a machine can do it.

The scale and ease of doing it matter.

It's like the difference between a bunch of shops having security cameras that they can look at the footage from, versus having every such camera connected to a wide surveillance network with facial recognition and querying abilities. Both are technically just a bunch of cameras in the same places, but not many people would argue it's the same thing.

Nope. Its the other way around. A human could not do it, but a machine can. Or at least it can fake that it can. And the trend seems to be to put the burden on the victim to prove the wrongdoing.

The AI bros who celebrate this development probably think the benefit they will extract by committing this on others exceeds the negatives of others committing it on them.

Good luck.

> You can’t easily prove that any outputs were generated by a particular model in general

But we also don’t know if there’s a trick up the sleeve that allows any of those models to be asked a very specific question that will result in a very specific answer.

Similarly to how map makers hide mistakes in their maps that they can point to when they want to prove that someone took their maps and presented it as their own.

Imagine using Llama for a customer support bot with a custom prompt, but there exist some phrase you don’t know that will make it say something that only Llama would say in response to that.

> But we also don’t know if there’s a trick up the sleeve that allows any of those models to be asked a very specific question that will result in a very specific answer.

So sanitize user input. Don't let the question get asked the specific way the user wanted it to.

Take a page out of OpenAI's book and sanitize the output too.

Matters a lot for any company trying to use them. There are licensing agreements, security audits, legal approval, etc. that make using any of these proprietary license models quite hard if not impossible

mistral appears to be quite open, and even better than llama imho

> 2023 was the year when "open source" got watered down to mean "you can look at the source code / weights"

Disagree, this has long been a problem, alongside all the other familiar deceptive-but-legal false advertising out there. Fortunately the HackerNews community knows enough to call it out. Last year I posted a big list of this happening on HN: https://news.ycombinator.com/item?id=31203209

I think it's time I rip off the GPL 3 in spirit and structure, but change everything into a license one issues himself called the Universal Pirates' License. It would be a document which sounds like a software license but in fact be a statement of reality: he who has and can copy the code gets to use the code, like a stick lying on the ground.

Check out our fully open recent 3b model which outperforms most 7b models and runs on an iPhone/cpu, fully open including data and details

Tuned versions outperform 13b vicuña, wizard etc

https://stability.wandb.io/stability-llm/stable-lm/reports/S...

Reply on news.ycombinator.com