Yann LeCun on why AI must be open source [video]
youtube.com
youtube.com
For clarity, llama etc are not open source because of the conditions on the license under which the weights are released. It may not be possible to copyright weights or to apply software licenses to them, it hasn't been decided by the courts. But everyone acts like it's possible, as evidenced by the licenses they use. So under that assumption, Meta is not using an open source license with the weights.
Edit: for greater clarity, the open source definition has the following clause
The license must not restrict anyone from making use of the program in a specific field of endeavor. For example, it may not restrict the program from being used in a business, or from being used for genetic research.
The Meta Llama license lists restrictions on field of endeavour. Ergo, not open source. There are other clauses (the one about the number of users) which would also disqualify it, but the use restrictions are the most egregious.
At the end of the day - having LLM tech that has an open source _code_ base is what many people are talking about, the "data" that forms the model isn't always open source and while it would be nice if that was the case - is somewhat similar to open source software vs the data you run in it.
Open source means something else.
Calling released trained models open source is akin to releasing the binaries of a software for free, without releasing the source code, and calling it open source.
If it was trained on the entire internet for example - you can't just bundled up every piece of data on the internet into a zip file to include with the code to interpret it.
To use something like Meta’s open sourced LLAMA2 model you don’t need the data. The model is self contained. It’s a compressed lossy form of all the data it was trained on.
The weights allow you to continue its training with new data of your choosing.
This is a straw man argument as Meta has released the model weights and not just the trained model. When it comes to models the weights are the source.
There's a philosophy of open source based on recipients being able to make any small and large modifications to the source material as they see fit.
Just being able to build new projects with something or read what it does not qualify in that context -- it requires the freedom to rework the thing itself to suit your own needs and ends.
In the case of these models, that philosophy would have the source training data and the training algorithms themselves available, editable, verifiable, and repeatable. Ignoring the financial challenges, a truly open source model would let you remove, add, edit, or reclassify a sample in the original training data and then build new weights of their won.
The released weights are not that kind of open source, and that's what these arguments are generally making the a case for.
There are still plenty of grounds from which to critique that argument if you wanted to, but there's no straw man involved in it.
The reason I see it as a strawman is because the weights are the source of the AI model and they do allow you to modify the model however you please.
The data is not owned by Meta, they can’t release it and it’s not required to use or modify the model. This is like wanting a team to release the research that informed product development with their source.
The data and training code are not the same as the AI model itself. This is not the same as code that produces a binary, this is more like code that produces another form of code.
If you try to replicate the model you won’t be able to because the process is not deterministic. There’s no reason to replicate the model, you would fine tune it instead.
I agree with the point you made elsewhere that releasing the data is irrelevant.
You demonstrably don’t understand what it means to open source an AI model. The code is not relevant at all, the same code will not produce the same model weights, it is not deterministic.
Also you will need millions of dollars to retrain, retraining is not the goal. The model weights are the source that was opened.
Model weights are not "the preferred form in which a programmer would modify the program" but more like "the output of a preprocessor or translator" [0], and thus are not "source code", but viewed as code, are more like object code. The training data, training configuration, and program source for the training routines are "source code", without which the model as a whole is not truly open source, just as much as the program source for the inference routines are.
[0] for the relevance of the quoted material to whether or not your description meets the Open Source Definition, see the definition itself: https://opensource.org/osd/
It's source available, but it's not "open source" in the historical usage by programmers. Microsoft could not take their model weights and use it. It's not open source.
The weights aren’t binary. It’s not a compile form, it’s the actual source of the AI model.
The same code used to train the model / generate the weights will not produce the same model weights if run again, it is not deterministic.
Also you will need millions of dollars to retrain, retraining is not the goal. The model weights are the source that was opened. The weights can be used to modify the model.
The weights are typically distributed as a binary blob containing IEEE floating point values, or fixed point values; and these numbers are typically not meant for direct human consumption.
> The same code used to train the model / generate the weights will not produce the same model weights if run again, it is not deterministic.
This is not true, except in contrived situations like when you use PRNGs and you deliberately throw away their seeds.
> Also you will need millions of dollars to retrain
This is besides the point.
Even I feel that "700M monthly active user" gives not-so-open vibes, but frankly many of these "open source" licenses have oddly restrictive requirements as well.
The inference and training software used to run the model are open source, the concrete model -- that is, the thing for which the weights are the object code -- is not.
The concrete model is free-to-use closed source, which is better than an undisclosed blob hiding behind a SaaS service, but still not open source.
It's also good that the inference and training code are open source, even though the training data and configuration is not.
I'm curious what you are drawing issue with, but it's also under the bridge. Any sane reading of this license and the release shows that you have WAY more rights compared to GPT-4 or other fully closed releases. The availability of these weights and code are propping up an open source community around LLMs, which lends oxygen to other even-more-open endeavors.
Open-source tends to foster monopoly and relies on free labor (see Google, Meta, Linux, etc.)
There is a weird definition of open source that creeps in here.
Llama2 is effectively, proprietary compiled software. GPT-4 is proprietary compiled and hosted as SaaS.
We have yet to see a true open source LLM...
AI/ML is reinventing so many terms it's a bit annoying.
It appears all my favorite books have been scanned by Open AI. I am assuming that they didn’t purchase them.
For AI, "open source" is less about the license and what it allows you to do legally, and more about what the technology allows you to build practically.
If/when Meta gets litigious, this might change in a hurry. Right now, there are bigger risks to AI startups (not Google) than worrying about how they'll choose to enforce the llama license.
Edit: Actually, if the training set is IP and the method to train is IP then maybe the weights are licensable, given that it's trade-secret territory. However, ensuring that the edge-weights are not widely disseminated once shared with a licensee is not straightforward and the infeasibility of such makes it an uphill climb. Not like that stops the lawyers from clawing up the tree, tho.
In this field where the metaphysical benefits (and risks) are so great, you simply cannot shade the truth. It's hard enough filtering everyone else's agenda on matters much less significant.
Am I the only one?
It's like finding out that everyone else has an entirely different concept in their head than the one I thought we agreed on a quarter-century ago.
*E.g., Ctrl+F "obscenity".
Firstly, I will state that I am in full agreement with LeCun on what he states in the video about the need for AI systems such as LLMs to be open [1]. I am not sure if I entirely agree with his very high optimism regarding AI safety, but I have always been far more concerned about current risks and I suspect that in practice we are very much aligned.
[1]: https://yewtu.be/watch?v=EGDG3hgPNp8&t=6525s
Now, let me move to the issue of licensing that I have already talked about at great length in comments on other, older submissions. So let me do something new here and try to put it into a historical context.
Computer science and machine learning in particular has a tradition of openness that goes beyond that of many other fields. What I mean by this is that over the last twenty or so years, there has been a gradual move towards open sharing of research, research outcomes, etc. All under the umbrella of open science and FLOSS. One of the earliest examples I am aware of (there may be others) is how the Journal of Machine Learning Research was established as a direct reaction to the resistance against open access in another leading journal at the time [2]. Similarly, in natural language processing, which my primary field of research, we have an open anthology where the great majority (if not nearly all) of the most important research going all the way back to the 70s is available for everyone to read and use [3]. In AI, we really have pushed back against the big publishers and this is a beautiful thing indeed (one which LeCun has contributed towards at that). So, given the context of this, I believe you can understand the dismay which the field feels when Google Deep Mind publishes papers such as the one on Alpha Go in journals without open access [4]. Yes, getting hold of the papers is not challenging for us, but it clearly goes against the efforts of the great majority of the filed. Thus the dismay.
[2]: https://en.wikipedia.org/wiki/Journal_of_Machine_Learning_Re...
[4]: https://www.nature.com/articles/nature24270
Moving now to FLOSS. When I entered the field nearly 15 years ago, it was somewhat of a rarity to find source code attached to papers. Likewise, training data was also not always made available or came under awful (and expensive) licensing agreements such as those imposed by the Linguistic Data Consortium. However, we as researchers released more and more open data and code, and by 2015 or so it was starting to become the norm that both code and data was attached to papers. Back in those days, training a system given the training data took maybe a day or so, thus model weights were less important. In addition, the models were less complex and thus replicating the training procedure was less problematic. Still, we saw more and more papers being published with their weights. In terms of licenses, there was a healthy mix of MIT-like, Apache 2.0, and GPL. The notable exception in the field would be Google Deep Mind, which despite their prominence at best would publish a library or two. This in contrast to say Google "Proper", which gave us say BERT [5]. It is in this context which I and many others view the last few years of closing down and adding restrictions to the sharing of model code and weights.
[5]: https://aclanthology.org/N19-1423
On the topic of Facebook, I have always considered Facebook AI Research (FAIR) to be the most open of the commercial research labs. For the first few years what FAIR seem to have defaulted to the same open standards as most of the research community. As an example, RoBERTA [6] was shared freely and was the basis for a lot of research for the next few years. However, at some point something changed inside FAIR and instead of releasing model code and weights under permissive licenses, they first changed their default license to exclude commercial usage, and now most recently ship with custom weights-available licenses and usage policies. This is still better than a lot of their competition, but I hope that given the historical context I have provided here you can see why many members of the community see this as the tide turning in terms of openness in the field and we are concerned. It should also be noted that FAIR still uses open science and open source as terms, despite not honouring the history of the terms.
[6]: https://arxiv.org/abs/1907.11692
I want to state that I am not of the position that we are entitled to the work of others. However, I believe the actions of FAIR and others have and will have negative consequences for the long term openness of the field and thus harm both scientific inquiry and industrial innovation. I also believe that it is blatantly dishonest of FAIR and others to claim to support open science and open source, while being in direct violation of this through their actions [7].
[7]: https://blog.opensource.org/metas-llama-2-license-is-not-ope...
A very fair line of reasoning states that what weights-available is still better than what "Open"AI and others are doing. I think this is a good point and I agree with it. Likewise, one could argue that maybe this is the best commercial actors can do in light of the current commercialisation and likely internal political pressure. I again think this is a good point, but I still feel that given the historical context the behaviour is still undesirable. Furthermore, I have heard through colleagues how there is now internal pressure inside several organisations to favour carrying out experiments on their models over those of the competition and I do not believe I need to tell you why this is problematic from a scientific perspective.
To finish this already lengthy post, allow me to highlight some problematic portions of the LLaMa 2 license [8].
> You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Llama 2 or derivative works thereof).
This explicitly forbids an ongoing line of research; thus puts limits on the scientific inquiries one can pursue.
> Additional Commercial Terms. If, on the Llama 2 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights.
This is a direct violation of both how the Free Software Foundation and Open Source Initiative defines FLOSS. It also puts up barriers for researchers at private research labs; thus has a chilling effect on the amount of research conducted.
[8]: https://ai.meta.com/llama/license
Apologies for the wall of text. Hopefully some will find some value in my ramblings, just like I have found in yours. I really should get around to writing a weblog...