Hacker News Reader

Comment by thrtythreeforty | Hacker News Reader

I'm very glad people are starting to push back against claims of various LLMs being open source. I was beginning to be worried that the term would be forcefully redefined in the ML space to mean "weights available." With the kickoff of projects like this and Databricks' Dolly, I'm heartened to see the community saying "no, we are willing to spend the compute to make actually open models."

(While it's true that the actual model code of Llama is properly open source, it's also useless for inference by itself. Claiming these models are open source seems like having your cake and eating it too - you get accolades for "open sourcing" but still get to control what happens with it.)

View on HN

I can only agree. The number of times we have seen corporations abuse “open source” and “open science” in the context of large language models have been baffling: OPT/LLaMA disallowing commercial usage, BLOOM having an ethical non-open license, GLM having a clause not to “undermine [the People’s Republic of China’s] national security and national unity”, etc. Every single one of these models have been happy to ride on the coattails of the hard work of the open movements by calling themselves open, while only paying lip service to the ideals and definitions underpinning them.

While RedPajama has yet to commit to a license (from what I can see, it is late at night…), they are making all the right noises and I am hopeful that my prediction that we are about to see the floodgates of truly open models blow open and that OpenAI’s “moat” will be proving to be a lot shallower than what they and many others have made us believe over the last six months will come true.

> not to undermine the national security and national unity

this is a required statement to conform with China’s constitution, or the superseding authoritative social contract there.

think of it like if the Patriot Act was an article of the constitution instead of a random law subservient to the constitution, it would negate other parts of the constitution that we hold near and dear.

this is a useful similarity as both constitutions have assurances of free speech

just one has a fatal heavily leveraged clause that undermines all other parts of that constitution and dictates all facets of life

This is interesting, thank you. But then how can any entity in the PRC contribute to open source? Alibaba, Baidu, etc. have released plenty of machine learning code under proper open licenses in the past (not to mention that we have hardware vendors in the PRC contributing to say Linux). The story I heard about GLM was that they were a high enough public profile project that it caught the attention of PRC bureaucrats that pushed for the clause to be included.

Regardless of the cause though, the clause flies afoul of any definition of open out there.

simplest answer is that Alibaba and Baidu have more party members as stakeholders

but its not likely that any uncontrollable LLM can start spitting out accuracy or things unhelpful to Beijing’s ethos there and be allowed to operate

the model or the service filtering the model has to be controlled

> this is a required statement to conform with China’s constitution

But doesn't this mean the model training data also excludes anything critical of China?

For example, does their training data include things like this: https://en.wikipedia.org/wiki/1989_Tiananmen_Square_protests... ?

go test it out and let us know, give it a really hard conversation like if beijing is sensitive to anything else from the last 34 years

Hi, this is Vipul, I am a co-founder of Together. We plan to release the model weights under Apache 2.0. The amount of creativity that Stable Diffusion unleashed for instance is only really possible with permissive licenses!

Are you working at all with Stability, Eleuther, or LAION? There have been some rumors that they are doing something similar to this and I'm wondering if this is a duplicated effort.

Either way, huge fan, it would be awesome to have a LLaMA set of weights that are fully open.

“Acknowledgements

We are appreciative to the work done by the growing open-source AI community that made this project possible.

That includes:

    Participants in building the RedPajama dataset including […] LAION.  

    Meta AI — […]. 

    EleutherAI — This project is built on the backs of the great team at EleutherAI — including the source code they provided for training GPT-NeoX. 

    An award of computer time was provided by the INCITE program. This research also used resources of the Oak Ridge Leadership Computing Facility (OLCF), which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725.”

The answer to your question is right there at the bottom of the page in the linked-to blog post :/

Thank you Vipul, you and the others are really doing god’s work and have the full support of myself and my academic research team, who are eager to push the boundaries with data, prompts, and investigations of whatever you release (in fact, we have spent the last couple of months working to produce multi-lingual prompts and enriching the few open models we had so far). Just a very quick point of feedback.

While I am not a lawyer and Apache 2.0 is likely to be unproblematic, I always find it puzzling as to why people recently are opting to license non-software using software licenses (Apache 2.0 in particular). Hopefully you have access to sensible lawyers, but I was always under the expectation that model weights would fall under a license such as CC-BY rather than Apache 2.0. Sadly it has been too long since I read the recommendations and justifications for this, so I can not find a good reference, but seem to recall the advice came out of FSF.

What about the training data/code?

https://salsa.debian.org/deeplearning-team/ml-policy

My only caveat here is that I'm actually really curious to see a ruling about whether model weights can be copyrighted.

I don't think the "Open Source" label people are using is accurate, and I heavily agree that a common thing that companies seem to be trying to do in this space is release what are essentially closed models while calling them open, and it's a really dangerous direction for AI to go. So nothing in your comment is wrong.

But it also feels a little bit like ceding ground to just assume that Llama can't be used commercially just because Facebook says it can't. I never signed a EULA with them, that claim depends entirely on whether or not model weights are under copyright (or under some similar form of IP protection, some people have brought up trade secrets).

And I don't have a super-strong opinion necessarily, but I'm not sure that's a safe assumption for people to make, and I kind of think it might be good to throw an asterisk next to "can't be used for commercial projects" whenever we talk about Llama's restrictions.

But again, I agree with you, it's not the same as saying Llama is Open Source. Even if it does get ruled as having weaker protections, I don't think the term would really apply.

I haven't done so, but don't you sign an agreement when you ask Facebook for a link to download the weights for LLAMA which is currently the only officially supported way of getting those weights (https://github.com/facebookresearch/llama/tree/main#llama) ?

Why would you bother using an "officially supported" way of downloading the weights if they aren't copyrightable anyway?

I haven't used Llama for anything other than playing around to test its capabilities, so I feel fairly comfortable admitting publicly that when I did that testing, I did not download it from Facebook using an official portal, and I didn't sign any agreement about it.

On that subject, to the best of my knowledge, I also haven't signed any kind of agreement with OpenAI. I've done all of my GPT testing through 3rd-party services or portals that don't require signing EULAs to use.

I got the weights via BitTorrent, so no I didn't sign/agree to anything.

To make an analogy with Linux, the weights are (up until now) a very large closed source firmware blob.

I like Debian's ML definitions, a "only weights available under libre license" situation is a "ToxicCandy" model. For a truly libre model you have to have libre GPU drivers/firmware, libre training data, libre training code, libre trained models and libre code to get outputs from the model.

https://salsa.debian.org/deeplearning-team/ml-policy

Lawyer here, still trying to wrap my head around all of it -- but it seems as if what may be different here is the extent to which all of this is practically "open-source" or even "literally free, as in freedom and cost etc" (i.e. generally and widely available REGARDLESS of what the law says)

And then coming second appears to be "companies and whoever who seek to make money, and intend to make some sort of legal restriction part of the biz model."

I have no answers or even predictions here except "this is gonna be interesting."

Reply on news.ycombinator.com