I'm just a lowly outsider to the AI space, but calling these open source models seems kind of like calling a compiled binary open source.

If you don't have a way to replicate what they did to create the model, it seems more like freeware than open source.

As an ML researcher, I agree. Meta doesn't include adequate information to replicate the models, and from the perspective of fundamental research, the interest that big tech companies have taken in this field has been a significant impediment to independent researchers, despite the fact that they are undeniably producing groundbreaking results in many respects, due to this fundamental lack of openness

This should also make everyone very skeptical of any claim they are making, from benchmark results to the legalities involved in their training process to the prospect of future progress on these models. Without being able to vet their results against the same datasets they're using, there is no way to verify what they're saying, and the credulity that otherwise smart people have been exhibiting in this space has been baffling to me

As a developer, if you have a working Llama model, including the source code and weights, and it's crucial for something you're building or have already built, it's still fundamentally a good thing that Meta isn't gating it behind an API and if they went away tomorrow, you could still use, self-host, retrain, and study the models

Which option would be better?

A) Release the data, and if it ends up causing a privacy scandal, at least you can actually call it open this time.

B) Neuter the dataset, and the model

All I ever see in these threads is a lot of whining and no viable alternative solutions (I’m fine with the idea of it being a hard problem, but when I see this attitude from “researchers” it makes me less optimistic about the future)

> and the credulity that otherwise smart people have been exhibiting in this space has been baffling to me

Remove the “otherwise” and you’re halfway to understanding your error.

> Release the data, and if it ends up causing a privacy scandal...

We can't prove that a model like llama will never produce a segment of its training data set verbatim.

Any potential privacy scandal is already in motion.

My cynical assumption is that Meta knows that competitors like OpenAI have PR-bombs in their trained model and therefore would never opensource the weights.

This isn't a dilemma at all. If Facebook can't release data it trains on because it would compromise user privacy, it is already a significant privacy violation that should be a scandal, and if it would prompt some regulatory or legislative remedies against Facebook for them to release the data, it should do the same for releasing the trained model, even through an API. The only reason people don't think about it this way is that public awareness of how these technologies work isn't pervasive enough for the general public to think it through, and it's hard to prove definitively. Basically, if this is Facebook's position, it's saying that the release of the model already constitutes a violation of user privacy, but they're betting no one will catch them

If the company wants to help research, it should full-throatedly endorse the position that it doesn't consider it a violation of privacy to train on the data it does, and release it so that it can be useful for research. If the company thinks it's safeguarding user privacy, it shouldn't be training models on data it considers private and then using them in public-facing ways at all

As it stands, Facebook seems to take the position that it wants to help the development of software built on models like Llama, but not really the fundamental research that goes into building those models in the same way

> If Facebook can't release data it trains on because it would compromise user privacy, it is already a significant privacy violation that should be a scandal

Thousands of entities would scramble to sue Facebook over any released dataset no matter what the privacy implications of the dataset are.

It's just not worth it in any world. I believe you are not thinking of this problem from the view of the PM or VPs that would actually have to approve this: if I were a VP and I was 99% confident that the dataset had no privacy implications, I still wouldn't release it. Just not worth the inevitable long, drawn out lawsuits from people and regulators trying to get their pound of flesh.

I feel the world is too hostile to big tech and AI to enable something like this. So, unless we want to kill AGI development in the cradle, this is what we get - and we can thank modern populist techno-pessimism for cultivating this environment.

Translation: "we train our data on private user data and copyrighted material so of course we cannot disclose any of our datasets or we'll be sued into oblivion"

There's no AGI development in the cradle. And the world isn't "hostile". The world is increasingly tired of predatory behavior by supranational corporations

This post demonstrates a willful ignorance of the factors driving so-called "populist techno-pessimism" and I'm sure every time a member of the public is exposed to someone talking like this, their "techno-pessimism" is galvanized

The ire people have toward tech companies right now is, like most ire, perhaps in places overreaching. But it is mostly justified by the real actions of tech companies, and facebook has done more to deserve it than most. The thought process you just described sounds like an accurate prediction of the mindset and culture of a VP within Facebook, and I'd like you to reflect on it for a sec. Basically, you rightly point out that the org releasing what data they have would likely invite lawsuits, and then you proceeded to do some kind of insane offscreen mental gymnastics that allow this reality to mean nothing to you but that the unwashed masses irrationally hate the company for some unknowable reason

Like you're talking about a company that has spent the last decade buying competitors to maintain an insane amount of control over billions of users' access to their friends, feeding them an increasingly degraded and invasive channel of information that also from time to time runs nonconsensual social experiments on them, and following even people who didn't opt in around the internet through shady analytics plugins in order to sell dossiers of information on them to whoever will pay. What do you think it is? Are people just jealous of their success, or might they have some legit grievances that may cause them to distrust and maybe even loathe such an entity? It is hard for me to believe Facebook has a dataset large enough to train a current-gen LLM that wouldn't also feel, viscerally, to many, like a privacy violation. Whether any party that felt this way could actually win a lawsuit is questionable though, as the US doesn't really have signficant privacy laws, and this is partially due to extensive collaboration with, and lobbying by, Facebook and other tech companies who do mass-surveillance of this kind

I remember a movie called Das Leben der Anderen (2006) (Officially translated as "the lives of others") which got accolades for how it could make people who hadn't experienced it feel how unsettling the surveillance state of East Germany was, and now your average American is more comprehensively surveilled than the Stasi could have imagined, and this is in large part due to companies like facebook

Frankly, I'm not an AGI doomer, but if the capabilities of near-future AI systems are even in the vague ballpark of the (fairly unfounded) claims the American tech monopolies make about them, it would be an unprecedented disaster on a global scale if those companies got there first, so inasmuch as we view "AGI research" as something that's inevitably going to hit milestones in corporate labs with secretive datasets, I think we should absolutely kill it to whatever degree is possible, and that's as someone who truly, deeply believes that AI research has been beneficial to humanity and could continue to become moreso

> I feel the world is too hostile to big tech

Lmao what? If the world were sane and hostile to big tech, we would've nuked them all years ago for all the bullshit they pulled and continue to pull. Big tech has politicians in their pockets, but thankfully the "populist techno-pessimist" (read: normal people who are sick of billionaires exploiting the entire planet) are finally starting to turn their opinions, albeit slowly.

If we lived in a sane world Cambridge Analytica would've been the death knell of Facebook and all of the people involved with it. But we instead live in a world where psychopathic pieces of shit like Zucc get away with it, because they can just buy off any politician who knocks on their doors.

> normal people who are sick of billionaires exploiting the entire planet

Don't understand what big tech does for humanity and how much they rely on it in the day to day. Literally all of their modern conveniences are enabled by big tech.

Rather dismissive particularly as Crowdstrike has laid a good chunk of that bare.

In my experience many 'normal people' understand far more than you deign credit, many are able to forgo modern 'conveniences' if pressed.

Crowdstrike merely shows how much people depend on big tech and they don't even realize how much they rely on it.

I think you have too much faith in the average person. They scarcely understand how nearly everything in their life has been manufactured on or designed on something powered by big tech.

The model is public, so you can at least verify their benchmark claims.

Generally speaking, no. An important part of a lot of benchmarks in ML research is generalization. What this means is that it's often a lot easier to get a machine learning model to memorize the test cases in a benchmark than it is to train it to perform a general capability the benchmark is trying to test for. For that reason, the dataset is important, as if it includes the benchmark test cases in some way, it invalidates the test

When AI research was still mostly academic, I'm sure a lot of people still cheated, but there was somewhat less incentive to, and norms like publishing datasets made it easier to verify claims made in research papers. In a world where people don't, and there's significant financial incentive to lie, I just kind of assume they're lying

> If you don't have a way to replicate what they did to create the model, it seems more like freeware

Isn't that a bit like arguing that a linux kernel driver isn't open source if I just give you a bunch of GPL-licensed source code that speaks to my device, but no documentation how my device works? If you take away the source code you have no way to recreate it. But so far that never caused anyone to call the code not open-source. The closest is the whole GPL3 Tivoization debate and that was very divisive.

The heart of the issue is that open source is kind of hard to define for anything that isn't software. As a proxy we could look at Stallman's free software definition. Free software shares a common history with open source and in most open source software is free/libre, and the other way around, so this might be a useful proxy.

So checking the four software freedoms:

- The freedom to run the program as you wish, for any purpose: For most purposes. There's that 700M user restriction, also Meta forbids breaking the law and requires you to follow their acceptable use policy.

- The freedom to study how the program works, and change it so it does your computing as you wish: yes. You can change it by fine tuning it, and the weights allow you to figure out how it works. At least as well as anyone knows how any large neural network works, but it's not like Meta is keeping something from you here

- The freedom to redistribute copies so you can help your neighbor: Allowed, no real asterisks

- The freedom to distribute copies of your modified versions to others: Yes

So is it Free Software™? Not really, but it is pretty close.

The model is "open-source" for the purpose of software engineering, and it's "closed data" for the purpose of AI research. These are separate issues and it's not necessary to conflate them under one term

> it seems more like freeware than open source.

What would you have them do instead? Specifically?

Release the training set and the code that was used to train the model, or stop calling it open source.

If you can't fork it and take the project in your own direction, it's not open source.

A good point.

Forgive me, I am AI naive, is there some way to harness Llama to train ones own actually-open AI?

Kinda. Since you can self-host the model on a linux machine, there's no meaningful way for them to prevent you from having the trained weights. You can use this to bootstrap other models, or retrain on your own datasets, or fine-tune from the starting point of the currently-working model. What you can't do is be sure what they trained it on

How open is it really though? If you're starting from their weights, do you actually have legal permission to use derived models for commercial purposes? If it turns out that Meta used datasets they didn't have licenses to use in order to generate the model, then you might be in a big heap of mess.

From a legal perspective, yea. If we end up having any legal protection against training AI models, legal liability will be a huge mess for everyone involved. From an engineering perspective, if all you need is the pretrained weights, there's not a clear way Facebook could show up and break your product from a technological perspective, as compared to if the thing is relying on, say, an OpenAI API key rather than a self-hosted Llama instance

I could be wrong but most “model” licenses prohibit the use of the models to improve other models

That's a good point. I expect it is ultimately unenforceable though. I'm describing training a model for myself, not for sale or public consumption.

They actually did open source the infrastructure library they developed. They don't open source the data but they describe how they gathered/filtered it.