nope. I am self-hosting. support is pretty good actually. llama.cpp supports it (v1.6 too; and in openai API server as well). ollama supports it. open-web-ui chat too.
using it now on desktop (I am in China, so no OpenAI here) and in cloud cluster on project.
Locally it's actually quite easy to setup. I've made an app https://recurse.chat/ which supports Llava 1.6. It takes a zero-config approach so you can just start chatting and the app downloads the model for you.
It is! Just downloaded it the other day and while far from perfect it's pretty neat. I uploaded a Gene Wilder/Charlie in the Chocolate Factory meme and it incorrectly told me that it was Johnny Depp. Close I guess! I run LLAVA and llama (among other models) using https://ollama.com
As a "web builder" I do think these tools will be very useful for accessibility (eventually), specifically generating descriptive alt tags for images.
Now seriously, by Llama being "sort of" open source, it does not seem to be something someone can fork and develop/evolve it without Meta, right? If one day Meta comes and says "we are closing Llama and evolving it in a proprietary mode from now on" would this Llama indie scene continue to exist?
If this is the case wouldn't this be considered a dumping strategy by Meta, to cut revenue streams from other platforms (Gemini/OpenAI/Anthropic) and contain their growth?
Realistically the problem here might just be that the concept of open source doesn't really fit machine learning models very well, and we should stop trying to force it.
Sharing the end product, but not the tools and resources used to produce it, is how open source has always worked. If I develop software for a commercial operating system using a commercial toolchain, and distribute the source code under GPL, we would call that software open source. Others who get the code don't automatically get the ability to develop it themselves, but that's kind of beside the point. I don't have the rights to publicly redistribute those tools, anyway; the only part I can put under an open source license is the part for which I have copyright.
Training data for a LLM like Llama works similarly when it comes to copyright law. They don't own copyright and/or redistribution rights for all of it, so they can't make it open, even if they want to.
If that seems unsatisfying, that's because it is. Unfortunately, though, I don't think the Free Software community is going to get very far by continuing to try to fight today's openness and digital sovereignty battles using tactics and doctrine that were developed in the 20th century.
It does fit it. Perfectly. It's incredible. Like an Internet of all Human Knowledge released before 1965. OpenAI could of done this. The battle to me is just people respecting ideas instead of saying they are impossible or unnecessary because what we have is good enough.
"good enough" is incredibly subjective here. Maybe good enough for you, but there are many things that are not possible with either the dataset or the weights being available.
And some things are impossible even with both the dataset and weights. Say you wanted to train the same model as is released, using Meta's hypothetically released training data. You also need to know the starting parameters, the specific hardware and it's quirks during training, the order the data is trained in as well as any other preprocessing techniques used to treat the text.
Considering how ludicrously expensive it would be to even attempt a ground-up retrain (as well as how it might be impossible), weights are enough for 99% of people.
Good-enough? Please please type out what a truely open source ai model with open weights and open data would be like. I picture it like a Tower of Babel! Very far from "Good-enough"!
Nice marketing that is written for investors. Let us translate:
> By making our Llama models openly available we’ve seen a vibrant and diverse AI ecosystem come to life [...]
They all use the same model and the same transformer algorithm. The model has an EULA, you need to apply for downloading it, the training data set and the training software are closed.
> Open source promotes a more competitive ecosystem that’s good for consumers, good for companies (including Meta), and ultimately good for the world.
So the "competitive" system means that everyone uses LLama and PyTorch.
> In addition to Amazon Web Services (AWS) and Microsoft’s Azure, we’ve partnered with Databricks, Dell, Google Cloud, Groq, NVIDIA, IBM watsonx, Scale AI, Snowflake, and others to better help developers unlock the full potential of our models.
>They all use the same model and the same transformer algorithm. The model has an EULA, you need to apply for downloading it, the training data set and the training software are closed.
Everything in that sentence is false except the training data part.
>So the "competitive" system means that everyone uses LLama and PyTorch.
This sentence shows you don't understand the LLM landscape and it's also false.
>Sounds really open
Correct. They partner with practically every vendor available for inference, which, isn't even needed if you run their models locally.
Meta has done a lot of wrong things over the years. How they are approaching LLMs is not one of them.
Could you provide a link for downloading the complete and exact training software for the latest models?
You need to provide an email address and click a license agreement. Then you get a download link that expires after a day. I do not have to do this with the Linux kernel. Perhaps you are downloading from within Meta and are not exposed to these issues?
"To download the model weights and tokenizer, please visit the Meta Llama website and accept our License.
Once your request is approved, you will receive a signed URL over email. Then, run the download.sh script, passing the URL provided when prompted to start the download.
Pre-requisites: Ensure you have wget and md5sum installed. Then run the script: ./download.sh.
Remember that the links expire after 24 hours and a certain amount of downloads. You can always re-request a link if you start seeing errors such as 403: Forbidden."
You never see that agreement with `ollama run`. It's not even a shrink-wrap licence - there's no indication there's a restriction at all between requesting the model and the API receiving requests for it. This situation is probably going to end up with the ollama folks getting a slap on the wrist and told to implement some shrink-wrap wording but until then, nobody can be bound by that licence because Meta can't demonstrate that anyone has seen the offer.
I think this is just the ollama site rehosting in violation of the license (unless there is some fine print I am missing). Huggingface makes you login and accept the agreement.
Meta has spent massive sums of money to train these models and they've released the models to the public. You can fine-tune the models. You can see the source code and the architecture of the model. The EULA is commercially-friendly.
You are free to quibble over how truly "open source" these models are, but I am very thankful that Meta has released them.
Open source developers have spent far more time to develop the truly free stack that Meta uses to power its business in the first place.
I am grateful to these developers. I am not grateful for a half open release and the redefinition of established terms. Which, judging by the downvoting in this thread, are now spread with fire and sword.
The stack was very usable in 2010. At that time, some gcc and kernel developers were employed by SuSE and RedHat. It was not common to be employed by a large corporation to work on open source.
Projects like Python were completely usable then. But the corporations came, infiltrated existing projects and added often useless things. Python is not much better now than in 2010.
So you have perhaps React and PyTorch. That is a tiny bit of the huge OSS stack. Does Meta pay for ncurses? for xterm? Of course not, it only supports flashy projects that are highly marketable and takes the rest for granted.
So no, only a tiny fraction of the really important OSS devs are employed by FAANG.
Should they? Both of those are client-side software that aren't even really being monetized or profited-off by Meta. You could maybe get mad at Meta's employees for not donating to the software they rely on, but in the case of ncurses and xterm they're both provided without cost. They're not even server-side software, much less a deliberate infrastructure decision.
There's an oddly extremist sect of people that seem to entirely misunderstand what GNU and Free software is. It does not exist to stop people from charging money for software. It does not exist to prevent private interests or corporations from contributing to projects. It does not exist to solicit donations from it's users. All of these are options that some GNU or FOSS projects can choose to embody, not a static rule that they must all abide by. Since Cathedral and the Bazaar was published, people have been scrutinizing different approaches to Free Software and contrasting their impacts. We don't have to champion one approach versus the other because they ultimately coexist and often end up stimulating FOSS development in the long run.
> Python is not much better now than in 2010.
C'mon, now. Next you're going to tell me about how great Perl is in 2024.
So, in this submission Meta adjacent opinions have called OSS supporters all sorts of names while being upvoted.
At least Meta is shows its true colors here. It must have hurt that the OSS position has arrived at the Economist yesterday, so everyone is circling the wagons.
Nobody here really has an agenda, least of all on HN where the majority of us hate Facebook like the living devil. Everyone remembers Cambridge Analytica and the ensuing drama, but we're also up-to-date on all of FAANG's exploits. Meta is a supporter of Open Source, and arguably contributes multitudes more than Apple or Amazon does. This idea that strings-attached weights releases tank their reputation is stupid; Meta's contribution is self-evident, and only looks stupid when you hold them to nonsense standards that no company would hold up to. Really, which Fortune 500 companies are donating to xterm and ncurses anyways? Is there anyone?
Again, there are arguments you can make that have weight but this isn't one of them. Every person with connection to wireless internet is running a firmware blob on their "open source" computer, it doesn't mean they're unable to bootstrap from source. Similarly, people that design Open Source infrastructure around Meta's binary weights aren't threatening their business at all. An "open" release of Llama wouldn't help those end-users, isn't even guaranteed to build Llama, and is too large to effectively fork or derive from. There's a good reason engineers aren't paying attention to the dramatic and insubstantial exposes that get written in finance rags.
Meta employs kernel developers (and MySQL developers and memcache developers and the people that created and released zstd and a lot more). Aside from all of these are also a bunch of python code developers, and you might want to recheck the performance improvements of 2010 vs 2024 python - much of it driven by FAANG developers!
A few decades ago an organization was founded specifically to address statements such as this one. That's why some early Microsoft attempts at competing with OS had to be called "shared source", not "open source".
Unfortunately they didn't do a competent job at addressing it, which would have involved trademarking the phrase. As a result, "Open Source" means whatever you, I, Meta, or anyone else wants it to mean.
This a hard stance if you talk just about the code (not the model weights). The llama community licence is a bit weird and probably not an OSI compliant licence, but close. Regarding weights this is different, but to me it is actually difficult to understand still now to aplly copyright law here. Having said that one nicht und erstand why certain stupid looking clauses went into the code licence. If we do not understand copyright of model weights and do not have court rulings in the use oft training data und er different copyright regimes (US and EU), I would not care too much. We are still in the Wild West.
“Close” is not good enough for using a term with a very specific meaning. OSI = open source, everything else is source-available (which its arguable that either even applies, because the source of the weights, the dataset, is not available).
I agree that for Llama, things are weird and they want to cover their bases, and that its better than nothing, but the specific use of “open source” is a long-running corporate dilution of what open source really means and I am tired of it.
Funny thing is that what they implemented in that licence is kind of a copy left thingy. Because you can use llama to train other models, but you would have to declare them derivative and thus keep the licence terms. (Not saying this is OSI compliant, but at least a IMHO a new kind of beast)
Having access to a weight doesn't make it open. Else you can make the argument that Microsoft Word is open source because you have access to the binary.
The access to modify (uptrain/finetune) these weights is the same between Meta & others, unlike with Word (where Microsoft has an advantage because they have code and can recompile it). I think this is the only thing which matters in practical terms.
I feel the arbitrary split between code and weights makes little sense when discussing if these models are "open source" in the copyleft meaning of the term. If the average user can't meaningfully use your product without agreeing to non-free terms then it's morally closed source.
Anything else and you're just open-source-washing your proprietary technology.
I tend to see weights as nothing more than data, data which may not even be copyrightable. But Meta keeps calling their data "open source" when they clearly do not release the model under an open source license, and that is terrible, awful and misleading.
Model Weights are not the Source. Why can't that be obviously like a binary isn't source code - a binary is compiled from source. You can open-license the data in a binary so it can be reverse-engineered / modded but that doesn't make it open source.
Llama did not license its training data. It’s almost impossible to prove a particular LLM was used to generate any particular text, and there’s likely a bunch of illegal content within the dataset used to train the model (as is the case for most other LLMs)…
So why should I care about following a license? They have no mechanism to enforce it. They have no mechanism to detect when it’s being violated. They themselves indicated hostilities to other licenses, so why not ignore it?
Again, how do you reliably prove that I used your model if I do a bunch of tricks to "hide" this?
i.e. high temperature, exotic samplers (i.e. typicality sampling), using lora/soft prompts/representation engineering, using it as part of a chain on top of other AI, etc
I don't care if every lawyer on earth is hired by Meta. Show me evidence that any particular LLM can be trivially fingerprinted based on its outputs. No, that "red token green token" paper on watermarking (https://arxiv.org/pdf/2301.10226) is not an example of this because of how trivial it is to defeat.
Edit: I can't reply to the comment saying "Subpoena" but this commentator seems to think that using any LLM at all is grounds for a court to issue a Subpoena requiring you to disclose which LLM you're using. If this actually happened, you'd see a massive chilling effect. Also, what stops someone from silently replacing the model with a non infringing one the moment someone starts asking questions?
I'm pretty sure that most courts aren't capable of getting expert testimony which is good enough to deduce that I silently swapped out my blarg_3.1 model which was made using a 1/3 llama3 merge with something else with gloop_1.5 which is no longer infringing.
Like seriously, I again ask, given the idea of courts with warrants and Subpoena's, why should I care about meta's licensing?
Edit2: If you're afraid of an employee leaking this info, don't tell your employees. Good thing clever model merging leaves no traces if you delete metadata!
If you scrape a large enough part of the internet, you're naturally going to get extremely illegal training data that you won't effectively filter out. I guarantee you that at least a tiny bit of highly classified information was not filtered out of most LLM training data (it wasn't found during the filter step), and it's quite remarkable that this and the above revelations have not led to anyone being Subpoena'd or related in regards to it.
So no, I think that folks will be literally the opposite of litigious on this issue. You want to play that game Zuck? Let's see what happens when I hire my researchers to find the dirt on your models dataset. We will see then who "settles out of court".
An employee (or a hacker) leaking the fact, maybe even leaking details, is all it takes to get the bloodhounds called on you and now you are subject to discovery. Sure, you can lie, but if you are doing anything large enough to get attention, you are exposing yourself to possible liability.
Sounds as if you might not have dealt with corporate litigation. There's no need for courts to issue subpoenas. Lawyers can issue them too, and for practically anything. If you don't respond they can issue a motion to compel in court, which if granted means you are legally required to respond with the requested documents. The real problem is that all of this starts to get very expensive if you want to fight it.
Meanwhile, all those lawyers need is to grab your text messages, emails, and slack messages. There's no need to look at code. If you have been up to something it's likely to come out pretty quickly.
p.s., Hiding your tracks after a subpoena is issued is a good way to end up in jail, at least in the US.
> Llama isn't open source at all. Stop using that phrase for your product featuring even an EULA
We don't have a commonly-accepted definition of what open source means for LLMs. Just negating Facebook's doesn't advance the discussion.
The open-source community is fractured between those who want it to mean weights available (Facebook); weights and transformer available; weights with no use restrictions (I think this is you); and weights, transformer and training data with no restrictions (obviously not workable, not even the OSI's proposed definition goes that far [1]).
In a world where the only LLMs are proprietary or Llama and the open-source community either remains fractured or chooses an unworkable ask, the latter wil define how the term is used.
We have that definition. The user needs the complete capability to reproduce what is distributed. Which means training data and the source used to train the model.
If you distribute the output of bison (say, foo.c) and not the foo.y sources, you would get pushback.
Then there is the EULA which makes it closed source right from the start.
> Which means training data and the source used to train the model
This makes an LLM (emphasis on large) that is open source per this definition legally impossible. In every jurisdiction of consequence.
People like the term open source. It will get used. It is currently undefined. If the choice is an impractical definition and a bad one, we'll get stuck with the bad one. (See: hacker v cracker, crypto(currency) v crypto(graphy), et cetera.)
You shouldn't redefine a term if it doesn't apply. Just make a new term. This is how we ended up with the MiB vs. MB confusion where some systems incorrectly use the latter, making it effectively useless because the reader doesn't know if they really meant MB or actually meant MiB instead.
Because it makes communication harder. The whole point of words is to communicate. If somebody releases a model that is actually open source, they would have no way to describe it. Then the bastardized meaning of open source could creep into other contexts until it is entirely meaningless.
Only in the short term. Then everyone gets used to the new status quo, and all is good, or even better because one doesn't have to make up entirely new terminology.
In the long term, it's better to have new terminology, so we can differentiate between open source and open weight models. Open weight is unambiguous, and the meaning is clear.
When you see a MIT-licensed repository, do you call a screenshot, icons or other image assets "open pictures"? That would be laughable. We don't need non-ambiguity outside of context.
Second, the purism argument is just silly. Is GCC not open source because there's no LaTeX for IA-32 manual, and its x86 machine instruction generator is clearly distilled from it?
The source for GCC is open. Open source does not mean that everything it depends on is also open source. Nobody uses open source in that way, and you are deliberately misusing the term in order to justify further misuse.
This is not an organic use of an altered meaning. The term is imposed by huge corporations who force it on their developers and everyone who wants to get funding in the Llama ecosystem.
It has nothing to do with natural language evolution.
> an organic use of an altered meaning. The term is imposed by huge corporations who force it
LLMs were entirely proprietary. In that context, Facebook put forward a foundational model one can run on their own hardware and called it open.
At the time, nobody had defined what an open-source LLM was. People came out to say Llama wasn't open. But nobody rigorously proposed a practical definition. The OSI has started, but they're being honest about it being a draft. In the meantime, people are organically discussing open versus proprietary models in a variety of contexts, most of which aren't particularly concerned with the OSI's definitions.
you are wildly conflating the difference between a naturally occurring evolution in a word's usage and meaning (ie: slang becomes canon becomes slang cycle), and intentionally misusing an existing, established meaning, and then pushing for that misuse (and misunderstanding) to become canon.
one is pragmatic, ergonomic, and motivated by advancing relationships between communicating persons in a naturally occurring way because the common denominator is a quick race to mutual understanding.
the other is manufactured, and not motivated by relation and advancing communication, but by how the shift in understanding benefits the one pushing for it, and often involves telling people how to think but is doing it through subversion.
> conflating the difference between a naturally occurring evolution in a word's usage and meaning (ie: slang becomes canon becomes slang cycle), and intentionally misusing an existing, established meaning
Sort of. I'm claiming the meaning of open source when applied to AI is unsettled. There are guiding principles that seem to imply Llama is not open source. But merely pointing that out without offering a practical alternative definition almost guarantees that the intentionally-misused definition Facebook is promulgating becomes the accepted one.
fair point, and i think i can appreciate more where you are coming from now.
however i do not think the alternatives need to be proposed at this moment in time because right now, the discussion is about holding people who intentionally reframe and misuse words accountable for their "double-speak" given the term's precedent.
conventionally, and by historical collective understanding, it is not open source.
i get you are attempting to highlight AI perhaps means this should be a definition reconsidered, but the irony here is the message itself is distorted due to the conflation, hence why consistency in language to me seems self-evident as a net good.
there is most certainly a difference between naturally occurring (which our brains reeeally support in terms of language development and symbolic communication), and manufactured (and therefore pushing for a word, or more aptly, a perspective's adoption).
i'd rather words manifest through a common need to reach mutual understanding as a means to relate to one another and this world, rather than having someone who stands to benefit from the change in definition, tell me what it means, and then expect me to just "agree", while they campaign around that and pretend it's the established definition (and not actually their own revised version).
it'd be one thing if people who were throwing the term around so loosely would be transparent: "Hey, we know this isn't historically what everyone means by OSS, but... that's OSS your OSS this is OSS, everything is OSS
instead a lot of these narratives are standing on the shoulders of the original definition and context of what it means to be OSS, and therefore the pedigree, implications (and whatever else for PR spin/influence), and simultaneously diluting what it means in the process as the definition gets further and further obfuscated by those influencing the change, and its pedigree is relied on as a distraction away from what is being done, or actually said.
It is not undefined. The wrong term is just repeated in marketing campaigns, by Meta developers and those building businesses on Llama until people believe it.
They could use the more correct Open Weights (which is still a euphemism because of the EULA).
But they do not, and they know perfectly well what they are doing. They are the ones responsible for these discussions, but they double down and blame the true OSS people.
That's fallacious. It's a problem of popular use (and sales incentives), not of definition.
"Open-source" represents a cluster of concepts but at the core of it there is a specific definition in spirit at least -- you can see the source for yourself, and compile it for yourself.
If the source is not available, why would you want to call it open-source? Just call it something else. As simple as that.
Definitions are very often proscriptive. Communication only works when people are able to understand the language being used. Imagine how well networks would function if we didn't have documented protocols that define what things mean and how they should be understood.
Nobody can "force" someone else to use the correct definitions of words, but when people disregard their established meanings they risk communication breaking down and the confusion and misunderstandings that follow. If I went around speaking nonsense or making up my own invented definitions for established words I shouldn't expect to be understood and others would be perfectly right to correct me or ask that I stick to using the well understood and documented meaning of words if I expect to have a productive conversation.
It's also perfectly fair to call out people who twist the meaning of words intentionally so that they can lie, mislead, and manipulate others. When it comes to products, companies can't just say "Words can mean anything I say they do! There are no rules!" to get away with false advertising.
> when people disregard their established meanings they risk communication breaking down
The meaning of words drifts in every living language.
> perfectly fair to call out people who twist the meaning of words intentionally
We don't have consensus around what open source means for LLMs. Facebook is pretending we do. But so is everyone in this thread claiming there is a single true definition of an open source LLM.
> The meaning of words drifts in every living language.
And it does result in a lot of confusion and misunderstanding until gradually people are taught the new definitions and how they are used. There are also groups of people who deliberately and continuously redefine words because they don't want to be widely understood. Some want to develop a means to signal to and identify others within their in-group, and some want to keep outsiders from understanding them so they can speak more openly in mixed company.
> We don't have consensus around what open source means for LLMs.
There are people who will argue about what open source means for anything. It's okay that open source means different things to different people, but it does result in confusion in discussions until people make their definitions clear.
I don't think that Facebook has earned the benefit of the doubt, in fact they've more than earned our skepticism, so it's very reasonable to see their new definition of "open source" as being nothing but marketing rhetoric at best, or at worst, as an attempt to twist our still developing consensus on what open source means into something that violates the philosophy/spirit of the open source movement.
Orwell was in large part against re-definitions of existing words and the removal of words in order to reduce the basis for productive thought.
Re-definition example: War is peace, freedom is slavery.
Since the new euphemism for downloadable models is a re-definition, Orwell would have been 100% against it. In fact the new use of "open source" is an Orwellian term.
i'm inclined to believe Orwell would have disagreed with OP, and would be asking himself -- why is there such a distinct push by those who benefit from the reframing, to reframe what Open Source means (compared to its already established meaning).
imo, this is patently not what Orwell documented and criticized via narrative example, and certainly was not what i took as his position on the evolution of naturally occurring languages, in the alluded to book -- he is a writer, and i imagine no doubt understands the importance of language and shared associations, and what it means for a language to naturally evolve its vernacular (accepted, common, or developing) -- through usage, or otherwise.
Orwell highlighted and warned against the consequences of people in influential positions of power intentionally distorting the collective associations with their new, updated versions of existing words, campaigning around those distortions, and intentionally reframing associations over time such that, the associations are polarizing and obfuscated, motivated by manipulation to benefit a select few, not motivated by advancing communication -- it certainly was not an example of society and its language naturally evolving "definitions" through usage.
and the novel wasn't a criticism against slang, or association/vernacular changing/evolving over time throughout collective use, nor was it a stance on requiring fixed, permanent, unwavering definitions -- it only emphasized how important it is to have consistent meaning.
he just wanted to encourage people to be skeptical of those pushing for the "different" or updated meaning of words, that clearly had a well-defined context, and association, previously -- why are they so dedicated and determined to "push" for a new meaning to get accepted, when there is a previously established and well accepted meaning already.
that doesn't sound natural to me, that sounds manufactured.
> The user needs the complete capability to reproduce what is distributed.
GPLv3 defines "source code" as the preferred form for making changes.
For most normal software that is identical to what you'd use to recreate it... but the way to make changes to an LLM isn't to rebuild it, but is to run fine-tuning on it.
Once again: for those who are new here. There is Free Software, which has a usefully strict definition.
And there is Open Source, the business-friendly -- but consequently looser -- other thing.
You can like one or both, they both have advantages and drawbacks.
But you cannot insist that "Open Source" has a very strict definition. It just doesn't. That's why the whole Free Software thing is needed, and IMHO, more important.
I agree except for your opinion in the end. But I know you're not alone in this opinion and it has been discussed to death. At this point it's more political than anything else in CS.
Open Source has every bit as strict of a definition as Free Software. Open Source as a term was coined and popularized by the OSI. The term may have occasionally been used in different contexts prior to the OSI, but it was never commonly applied to software before that.
One could argue that the OSI should have gotten a trademark on the term. But the FSF doesn't have a trademark on the term "free Software" either, so the terms have approximately equal legal protections.
Meta using the term "open source" to apply to their model data when their license isn't an open source license is dishonest at best.
I think I agree that they shouldn't use "open source," but again, this confusion highlights that you have to put in work when it comes to this topic.
Free Software has the GPL, and all its related, healthy controversy. It's not perfectly clear, but it's far more battle-tested than the much more nebulous "Open source."
People who like "free software" put in work, and better understood that, to some extent, you can't have your cake and eat it too. "Open Source" is much more about a whole lot (to me, naive) wishful thinking.
(The OSI is a bunch of companies in a trenchcoat, the creators of the GPL were more principled.)
The FSF accepts a lot more licenses than just the GPL family of licenses as free software. As for the OSI, I don't think they have ever hidden who they are, how the organization is ran, etc. https://opensource.org/about
I just noticed they are currently discussing what "Open Source AI" should mean. You can join in and add your thoughts tot he discussion.
Sure, but you can't give something away for free that you don't own. What people complaining about LLama not being open source are talking about is the training data, and that isn't something that Meta owns for the most part.
The OSI doesn't have a monopoly on the definition of the words "open source". There is "open source as per the OSI Open Source Definition" and there are other interpretations.
Probably will get flagged, but I get so annoyed by the cynical takes on Meta and their open source strategy. Meta is the only company releasing true open source (React, pytorch, graphql) and now LLama. This company has done more for software development than any other in the last decade. And now they are burning down the competition in AI, making it accessible to all. Meta software engineering compensation strategy pushed up the high end of developer compensation by almost twice. Enough with the weird cynicism on their licensing policy.
Look at Apple spending a billion on ads to say they respect your privacy or the Earth. Meta is buying / licensing a market sector in an industry they dominate where they have full control of our data. Our data is what got them that billion dollars.
if something requires an EULA it isn’t open at all, it is just publicly available. By your logic, public services are “open source.” There are myriad corporations that release actual open source software that is truly free to use. If you experience massive success with anything regarding Meta’s LLMs, they’re going to take a cut according to their EULA.
I'm trying to figure out the logic that makes "free for commercial use with less than 700 million monthly active users" less open than "free for non-commercial use", which is the traditional norm for non-copyleft open source machine learning products. But I just can't get there. Could somebody spell it out for me?
You’re certainly entitled to the opinion that an agreement (as in EULA) is distinct from a license (as in GPL, MIT etc).
But many legal minds close to this issue have moved to the position that there is no meaningful distinction, at least when it comes to licenses like GPL.
I don't think Facebook/Meta is the beacon of open-source goodness you think it is. The main reason they created yarn instead of iterating on npm is to use their own patent-friendly license they wanted to use with React (before the community flipped out and demanded they re-license it as MIT). Early Vue adoption seemed mostly driven by that React licensing fiasco.
Get real? How about get the point? It would not have (could not have) been so widely adopted without being MIT licensed and open source. It was huge news when Facebook attempted to change their licensing, and they issued an apology then made it MIT when the community responded. https://meitar.com/wp-content/uploads/2021/08/facebook_react...
There is nothing weirdly cynical about it. This is a fact of life in Silicon Valley - that a lot of FLOSS is released for strategic reasons (such as building up a community before enclosing it to extract a profit), and not because the Grinch's heart grew 2 sizes one day. "Commoditize your complement": https://gwern.net/complement
You can benefit a lot from it, and I have... but do be sure you know what you are ferrying on your back before you decide to offer it a ride across the river.
No. I think they use that web & ML software to help run their $1.3 trillion marketcap online social network company, on which, I am given to understand from US Congressional hearings, they sell ads.
> a lot of FLOSS is released for strategic reasons (such as building up a community before enclosing it
Not only is "a lot" of FOSS not released like this, both free software and Meta's models cannot be monetized post-release. If Meta decides to charge money for Llama4, then everyone with access to the prior models can keep their access and even finetune/redistribute their model. There is no strategic flip Meta can attempt here without shotgunning their own foot off.
It absolutely is released like that. Please note that 'such as' does not mean 'all', 'a majority', or anything like that. It simply means 'one example'. (I eat many things 'such as' honey, yet that doesn't make me Winnie the Pooh.) And enclosures come up regularly here on HN where suddenly a company has embarked on a new incredible journey involving 'not releasing all the source code anymore under an OSI-approved license', so they are not even that exotic an outcome.
Today, for example, on Twitter, people are pointing out that Runway, which made a big deal out of how they were releasing DL models and not keeping them proprietary, has quietly deleted all its stuff on HuggingFace. (Much like how, say, Mistral decided that it was no longer releasing all its models, after presenting itself as the big open model champion opposed to OA.) Only a five-year-old child is either surprised or angered by this turn of events.
Twitter is the epitome of people who "don't get it". For one, you don't use Twitter anymore because the damn website doesn't exist. Your new website, 'X' is filled with boiling frogs that (somehow) expect things to get better. People that don't get the proverbial memo when a neo-Nazi buys their website and replaces it's logo with a half-swastika. It would not be exaggeration to say that the quality and prestige of 'X' chatter is about as worthwhile as 4chan's resident opinion.
For two - they can't do shit. Every one of those businesses can try to pivot into being the next OpenAI, but even OpenAI can't be fucked to turn a profit. They can't retract their models (remember Llama 1.0? hahahahahahahahaha) and they can't piss and moan to the authorities when their weights leak because 99% of them contain unlicensed copyrighted material in the first place.
The real twist to this Gordian knot? Open source is the only model you can release conventional AI under, because anything else precludes a Fair Use defense. Everyone that has "proprietary" models only manages it by playing keep-away with the community.
> Only a five-year-old child is either surprised or angered by this turn of events.
Five year old children know that they can keep their toys once they stop selling them at Wal-mart. Pulling a model off HF doesn't stop new finetunes or even redistribution and re-quantization of the old model. Mistral and Runway can do whatever they want, all their models are belong to us.
True? The cynicism is arguments over "true". If you trick yourself into believing this is what Open Source looks like (no source data) then you lose out on imagining what a real open source AI with open source data would be like.
Sounds like God. What is private? What is public? There is a line between an omnipresent God that sees everything (zero privacy) and an Open Source AI that collects and indexes all public information. I want an AI to be programmed / trained on our conversation here. (if your reading this God / FB / OpenAI / 2025 ai Source Dataset <3)
Open source and open source model is not a term that came from dictionary, and it is just based on what community thinks it means. As long as open source model doesn't cause confusion, which it does not as open source model today just means open weights model, fighting over it is not worth it.
Here we go again with the co opting of open source and the marketing open washing. Llama isn’t open source. Sharing weights is like sharing a compiled program. Without visibility into the training data, curation / moderation decision, the training code, etc Llama could be doing anything and we wouldn’t know.
Also open source means the license used should be something standard not proprietary, without restrictions on how you can use it.
> Sharing weights is like sharing a compiled program.
Not at all. They're only similar in the sense that both are a build artifact.
> Without visibility into the training data, curation / moderation decision, the training code, etc Llama could be doing anything and we wouldn’t know.
"could be doing anything" is quite the tortured phrase, there. For one, model training is not deterministic and having the full training data would not yield a byte-perfect Llama retrain. For two, the released models are not turing-complete or filled with viruses; you can open the weights yourself and confirm they're static and harmless. For three, training code exists for all 3 Llama models and the reason nobody uses them is because it's prohibitively expensive to reproduce and has zero positive potential compared to finetuning what we have already.
> Also open source means the license used should be something standard not proprietary, without restrictions on how you can use it.
There are very much restrictions on redistribution for nearly every single Open Source license. Permissive licensing may not mean what you think it means.
I think that focusing primarily on the discussion of what is or isn't open source software makes us miss an interesting point here, that Llama enables users to have a similar performance to frontier models in your own systems, without having to send data to third-party sources.
My company is building an application for an university client, regarding the examination of research data written in "human language" (mostly notes and docs).
Due the high confidentiality of the subjects, as often they deal with non-patented information, we couldn't risk using frontier models, as it could break the novelty of the invention, therefore losing patentability.
Now with Llama3.1, we can simply run these models locally, on systems that is not even connected to the internet. LLMs are mostly good in examining massive amount of research papers and information, at least for the application we are aiming at, saving thousands of hours of tiresome (and very boring) human labour.
I am trying to endorse Meta or Zuckerberg or anything like that, but at least in this aspect, I think Llama being "open-source" is a very good aspect.
To me it's fairly interesting how relatively little money it takes meta to pose a risk to other models makers businesses, who are dependent on having to run the model after they created it (because that is how they make money) while meta does not even have to deal with the cost attached to providing inference infra, at all, to pose that risk.
Can you imagine how incredible an open source model would be for research / humanity beyond the buisness needs right in front of us?
Open-Knowledge source with an Open-Inteligence that can guide you through the entire massive digital library of its own brain. Semantic data light-years beyond a Search Engine.
No, I really can't imagine it. Extrapolating from our free commercially-licensed offerings it would seem most people would ignore it or share stories on Reddit about how FreeGPT poisoned their family when generating a potato salad recipe.
An open source model would be able to give you the sources of its potato salad recipe inspiration. It would be the best of both worlds. AI Knowledge + Real Open Human Knowledge.
I'm pretty sure Kagi is like a normal search engine with AI integration like Google. Not an AI designed to be open source with an open dataset of knowledge it was trained on.
> pretty sure Kagi is like a normal search engine with AI integration like Google
Sure. The point is the thing you said only an open-source model can do, it can do. Plenty of proprietary LLMs can cite sources.
The plain truth is most of the benefits of open models are not on the consumer side. (Or at least, I haven't seen any articulated.) They're on the producers'. Open models are better for those of us training models. That's partly why the open data debate is academic--very few people are training large foundation models because the compute and electricity costs are prohibitive.
I'm kinda hoping World Governments will use their Public Library infrastructure to train AI. Japan is my #1 hope with how they are opening public science knowledge. Super-computers have been prohibitive for a long time but national science institutions could be a great place for open source & open weight AI.
Just because you have the dataset doesn't mean you can generate a reference. Let's say I hand you a potato salad recipe and a copy of the entire internet. Say you somehow extract all potato salad recipes from the dataset (non trivial btw) and none of them are an exact match for the recipe the model generated. Now what?
> Open-Knowledge source with an Open-Inteligence that can guide you through the entire massive digital library of its own brain. Semantic data light-years beyond a Search Engine
This sounds like the usual AI marketing with the word "open" thrown in. It's not articulating something youc an only do with an open source LLM (and doesn't define what that means).
I'm personally not thrilled with how locked down LLMs are. But we'll need to do a better job at articulating (a) a definition and (b) the benefits of adhering to it versus the "you can run it on your own metal" definition Facebook is promulgating. Because a model meeting Facebook's definition has obvious benefits over proprietary models run on someone else's servers.
No. But you haven’t articulated why making everyone’s Facebook chats public is a net good. What does opening that data up confer in practical benefits?
Given what we know about LLMs, one trained only on public-domain data will underperform one trained on that plus proprietary data. If you want source data available, you have to either concede the "open" models will be structurally handicapped or that all data must be public.
I’m not sure what they’re talking about, but I’ll throw my hat into the ring. Copyright and other such systems are destroying any chance that we, as humanity, have of letting LLMs progress in an open and transparent manner. We have to hide the training data and make the weights a black box because of such antiquated notions such as copyright. While I am willing to permit some level of exclusivity with creative works, 100+ years is unreasonable and stagnates human creativity even outside of ML tasks. In the 19th century, I could take a book I was raised on and write my own fanfiction, and because that book would have been public domain by the time I was an adult I could add onto the work and the other fans of the previous work can build upon it with me. We see this with Sherlock Holmes for instance. If I wanted to publish a book set in the world of Harry Potter I’d need to wait for JK Rowling to croak, and then wait another 70 YEARS.
We need dramatic reforms on copyright, as we’ve really let corporate interests crowd out our rights to human culture and ideas. While I alone cannot decide what we as a country should find reasonable, I can say I find 20 years + 5 years extension is perfectly reasonable and that corporations should have never been able to pay off politicians to get what they wanted. Let alone Sonny Bono, that bastard, signing in bills that specifically benefited him.
So, to reiterate, the idea I feel that corporations want to destroy is the idea that we, as a people, have rights to the works that form our popular culture and that no one man, let alone a faceless corporation, should be able to profit from a singular work for hundreds of years.
If you have the model weights you have roughly the same opportunities as the company that trained the model. The code you need to run inference on the Llama weights is very much open source. The only thing you're missing out on is the training code, which is prohibitively expensive to run for most anyways. Open source training isn't going to give you any unique insights into the "digital brain library" of your model.
Also just to be clear, if you want to set up a RAG with an open weight model and a large dataset there's nothing stopping you. Download Red Pajama and Llama and give it a try.
You're not really asking for an open source model though, you're asking for open source training data set(s), which isn't something that Meta can give you. There are open source web scrapes such as The Pile, but much of the more specialized data needs to be licensed.
I'm asking for an "Open Source AI" and Meta and everyone supporting them is convinced its impossible in our lifetimes :( We are living in the Dark Ages where Information = $$$. I pray to AI we one day grow out of this pointless destructive economic spiral towards the heat death of the Earth and collect and share open knowledge across all human cultures and history.
Well, as long as by "AI" you are referring to pre-trained transformers, then what you are effectively asking for is the data used to pre-train them.
OTOH why you want the data is not clear. You don't need it to run Meta'a models for free, or to fine-tune them for your own needs. The only thing the data would allow you is to pre-train from scratch, in other words to obtain the exact same set of weights that Meta is giving you for free.
All of that data is already available, just look into “shadow libraries”. Now, I do wish Meta and other companies would publish their data sets and we, as humanity, could improve upon them and empower even better LLMs, but the unfortunate reality is copyright is holding us back. Most of what you say is essentially gibberish, but there is truth that LLMs would be better if it could not only utilize its weights, but reference and search its training data (that is collectively owned by humanity, by the way) and answer with that and not just what it “thinks”.
They are never going to stop saying this or show us the actual source data. Imagine if they did... Do they even entertain the idea? Can they really not imagine Open Source AI being possible because of all the personal data they train on?
Yoshua Bengio is a very respected scientist with well deserved reputation, but this discussion is upsetting… “academia now trains much smaller models… 10^26 FLOPs is 10 to the 26 floating point operations per second.. yes.. how big is that compared to GPT-4? It is much bigger than all the existing ones…” (flops has a different meaning: there is no per second in the law; one single H100 from last year performs 1e15 FLOPs per second; llama3.1 was close to the 1e26 limit this year, and the total training FLOPS of other models are not published; research could change once compute is even cheaper but state laws move at glacial speeds…).
It is disheartening so see the damage capacity in the hands of a couple of paranoic people who perhaps read the wrong scifi and had lots of power to influence others. If California passes this law, in a few years the world economy will be very different.
I think companies will simply quit doing business in California. They are killing the golden goose with all these regulations, just like they poison their own real estate markets allowing NIMBYs to dictate regulations that keep housing prices high and allow petty criminals to run rampant while cities like San Francisco continue to diminish and suffer
> I think companies will simply quit doing business in California
Sure they will. Just like how the auto manufacturers stopped trying to sell in CA after stricter emissions regulations and just like how companies stopped trying to serve European visitors after GDPR came into force. California is a $4T GDP, placing it just behind Japan and ahead of India in global rankings. I am sure they are happy to tell non-complaint companies to piss off and not let the door hit them on the way out. This is the canary in the coal mine for cowboy AI companies, once EU regulators get into the game their days of playing fast and loose with this tech will be over.
Unfortunately, it increasingly feels like the cowboy is the lobbyist in this picture and the victims of their scam are the regulators and people with semi religious beliefs. The specific limits in the law are possible to reach with this year’s hardware and enough money to buy a large gpu cluster and let it train for months. In five years the current limit in the law might be accessible by academia using government clusters and in ten years it may be accessible to patient home users, especially if some of the specialized hardwares that are now in the works find ways to commercialize and have mass appeal. Limiting the amount of compute by law is such an odd concept when we are in a phase of exponential growth.
Dont underestimate the glacial kinetics of existing laws. I dont recall many examples of legal regulation that took lots of lobbying and years of planning to pass, to then be changed within years. Sometimes non-permanent laws, such as executive orders, change when executives change, but laws are typically only created and stay until they are either deemed unconstitutional or completely out of sync with reality, and even then only if there is nothing more important or new to write laws about. The more common end of bad laws is for executives to decide to not enforce them, eg outdated federal or state laws about drugs, but it takes a lot of momentum and time to undo a bad law, much more time than to write it in the first place. Just like with bad code in large codebases, new code is easier to write and brings more glory than carefully eliminating or amending poor code.
IMHO it's not, it just parrots the same old arguments for "safety", arguing against straw-men and framing the other side as having wrong assumptions about AI safety/being unfair/etc, all while not going into the principled counter-arguments and their own assumptions at all.
Here are some counter-points:
Regulation:
- Very little effort is made to evaluate risk of over-regulating, regulatory capture and counterproductive wrong regulation
- The downside of under-regulating is vastly overemphasized, most arguments boil down to "we have to act FAST now or x BAD thing might happen"
- The risk of over/wrongly regulating is vastly under emphasized with the same FUD reasons.
- according to one of the many straw-men argument in the pod I'm a libertarian against any and all regulation because I criticize possible regulatory capture, I would enthusiastically support regulation that foundation models have to be:
-- given freely to public researchers/academics for in-depth independent safety-research
-- open weighted after a while (e.g. after ~ a year, safety concerns should be mostly ruled out and new generations are out so ROI is already likely there. [e.g. there's NO safety reason at all for ClosedAI to not release gpt-3.5, llama3 is better already])
Proposed FLOP cut-off of SB 1047:
- according to the pod, the cut off is much more advanced than anything currently released.
- The 10^26 FLOP cutoff is way to low, llama-405b is ~4×10^25 FLOPs
- 405B is maybe 20% smarter than 70B, while taking over an order of magnitude more FLOPS to train, the cutoff itself is very likely not much smarter than the current SOTA.
- IMO none of the current SOTA models are very dangerous, but kill switch regulation is.
Kill-Switches:
- SB 1047 is (non-explicitly) calling for kill-switches over the cut-off due to liability of the model creators and market dynamics
- Any kill-switch regulation means a complete dead-end to any advanced open-weights AI. This means that huge corporations and governments will control any and all advanced AI-development. This is top-down control of the maths you are allowed to run on your computer IMO that is Orwellian as fuck.
China:
- mentioning china is FUD 101, it's basically AI's "think of the children"
- If they think they can stop china from building their own advanced LLMs, they're delusional. This regulation might even help them to get there faster. They don't even need to steal, there maybe a year or two behind the SOTA and catching up fast.
I just don't get how so many people on a site with "hacker" in the name want to make it impossible to hack on these things for anyone not employed by the big corporate AI research labs.
I haven’t heard anything specific from Meta themselves, but I think the bill is short enough that we can reason as non-lawyers about it. Almost certainly they would have to stop releasing LLMs weights based on the very specific qualifications in the legislation. I don’t actually know what the specific size limit would be, but based on the translated $ value in the text of the bill it probably would cover their 70B+ models.
Disturbing approach to mitigating AI harms imo, this bill basically hopes it can limit the number of operators of an arbitrary model type so as to allow easier governance of AI model use. This ignores the reality that we already have large models openly released and easily modifiable (outside CA jurisdiction) which likely are capable of perpetuating “critical harms”, or that the information requirements to achieve the defined “critical harms” could be realized by an individual by simply reading a few books. There’s also no reason to simply assume that future models will require millions of dollars of compute to create; fulfilling the goals of this regulatory philosophy long term almost certainly requires the banning of general purpose compute to come close to the desired outcome of a supposed reduction in probability of some “critical harm” being perpetrated. We should be focusing on hardening society to the realities of the “critical harms” identified by this bill, rather than implicitly assuming the only reason we don’t see them as much irl is because everyone is stupid. The current paranoia wave around LLMs is just a symptom of people waking up to the fragility of the world we live in.
190 comments
[ 3.3 ms ] story [ 256 ms ] threadusing it now on desktop (I am in China, so no OpenAI here) and in cloud cluster on project.
Locally it's actually quite easy to setup. I've made an app https://recurse.chat/ which supports Llava 1.6. It takes a zero-config approach so you can just start chatting and the app downloads the model for you.
What's your reservation about running it locally?
As a "web builder" I do think these tools will be very useful for accessibility (eventually), specifically generating descriptive alt tags for images.
Now seriously, by Llama being "sort of" open source, it does not seem to be something someone can fork and develop/evolve it without Meta, right? If one day Meta comes and says "we are closing Llama and evolving it in a proprietary mode from now on" would this Llama indie scene continue to exist?
If this is the case wouldn't this be considered a dumping strategy by Meta, to cut revenue streams from other platforms (Gemini/OpenAI/Anthropic) and contain their growth?
Sharing the end product, but not the tools and resources used to produce it, is how open source has always worked. If I develop software for a commercial operating system using a commercial toolchain, and distribute the source code under GPL, we would call that software open source. Others who get the code don't automatically get the ability to develop it themselves, but that's kind of beside the point. I don't have the rights to publicly redistribute those tools, anyway; the only part I can put under an open source license is the part for which I have copyright.
Training data for a LLM like Llama works similarly when it comes to copyright law. They don't own copyright and/or redistribution rights for all of it, so they can't make it open, even if they want to.
If that seems unsatisfying, that's because it is. Unfortunately, though, I don't think the Free Software community is going to get very far by continuing to try to fight today's openness and digital sovereignty battles using tactics and doctrine that were developed in the 20th century.
Considering how ludicrously expensive it would be to even attempt a ground-up retrain (as well as how it might be impossible), weights are enough for 99% of people.
> By making our Llama models openly available we’ve seen a vibrant and diverse AI ecosystem come to life [...]
They all use the same model and the same transformer algorithm. The model has an EULA, you need to apply for downloading it, the training data set and the training software are closed.
> Open source promotes a more competitive ecosystem that’s good for consumers, good for companies (including Meta), and ultimately good for the world.
So the "competitive" system means that everyone uses LLama and PyTorch.
> In addition to Amazon Web Services (AWS) and Microsoft’s Azure, we’ve partnered with Databricks, Dell, Google Cloud, Groq, NVIDIA, IBM watsonx, Scale AI, Snowflake, and others to better help developers unlock the full potential of our models.
Sounds really open.
Everything in that sentence is false except the training data part.
>So the "competitive" system means that everyone uses LLama and PyTorch.
This sentence shows you don't understand the LLM landscape and it's also false.
>Sounds really open
Correct. They partner with practically every vendor available for inference, which, isn't even needed if you run their models locally.
Meta has done a lot of wrong things over the years. How they are approaching LLMs is not one of them.
You do need to apply on Huggingface to download the model.
> This sentence shows you don't understand the LLM landscape and it's also false.
PyTorch definitely is the most used ML framework.
You need to provide an email address and click a license agreement. Then you get a download link that expires after a day. I do not have to do this with the Linux kernel. Perhaps you are downloading from within Meta and are not exposed to these issues?
I am confused - I grabbed Ollama and pulled down some of these models. I don't recall having to go through any legal agreements. I just type:
Maybe I missed something and am actually 10 steps behind. Who knows anymore. This whole space is totally insane to me."To download the model weights and tokenizer, please visit the Meta Llama website and accept our License.
Once your request is approved, you will receive a signed URL over email. Then, run the download.sh script, passing the URL provided when prompted to start the download.
Pre-requisites: Ensure you have wget and md5sum installed. Then run the script: ./download.sh.
Remember that the links expire after 24 hours and a certain amount of downloads. You can always re-request a link if you start seeing errors such as 403: Forbidden."
https://ollama.com
You may come away surprised.
https://ollama.com/library/llama3.1/blobs/0ba8f0e314b4
> By clicking “I Accept” below or by using or distributing any portion or element of the Llama Materials, you agree to be bound by this Agreement.
I think my lawyer would have a few things to say about automatic legal agreements hidden somewhere in source control.
Is the ollama project part of Meta? Is that what's going on here?
You are free to quibble over how truly "open source" these models are, but I am very thankful that Meta has released them.
I am grateful to these developers. I am not grateful for a half open release and the redefinition of established terms. Which, judging by the downvoting in this thread, are now spread with fire and sword.
Projects like Python were completely usable then. But the corporations came, infiltrated existing projects and added often useless things. Python is not much better now than in 2010.
So you have perhaps React and PyTorch. That is a tiny bit of the huge OSS stack. Does Meta pay for ncurses? for xterm? Of course not, it only supports flashy projects that are highly marketable and takes the rest for granted.
So no, only a tiny fraction of the really important OSS devs are employed by FAANG.
Should they? Both of those are client-side software that aren't even really being monetized or profited-off by Meta. You could maybe get mad at Meta's employees for not donating to the software they rely on, but in the case of ncurses and xterm they're both provided without cost. They're not even server-side software, much less a deliberate infrastructure decision.
There's an oddly extremist sect of people that seem to entirely misunderstand what GNU and Free software is. It does not exist to stop people from charging money for software. It does not exist to prevent private interests or corporations from contributing to projects. It does not exist to solicit donations from it's users. All of these are options that some GNU or FOSS projects can choose to embody, not a static rule that they must all abide by. Since Cathedral and the Bazaar was published, people have been scrutinizing different approaches to Free Software and contrasting their impacts. We don't have to champion one approach versus the other because they ultimately coexist and often end up stimulating FOSS development in the long run.
> Python is not much better now than in 2010.
C'mon, now. Next you're going to tell me about how great Perl is in 2024.
At least Meta is shows its true colors here. It must have hurt that the OSS position has arrived at the Economist yesterday, so everyone is circling the wagons.
Again, there are arguments you can make that have weight but this isn't one of them. Every person with connection to wireless internet is running a firmware blob on their "open source" computer, it doesn't mean they're unable to bootstrap from source. Similarly, people that design Open Source infrastructure around Meta's binary weights aren't threatening their business at all. An "open" release of Llama wouldn't help those end-users, isn't even guaranteed to build Llama, and is too large to effectively fork or derive from. There's a good reason engineers aren't paying attention to the dramatic and insubstantial exposes that get written in finance rags.
I agree that for Llama, things are weird and they want to cover their bases, and that its better than nothing, but the specific use of “open source” is a long-running corporate dilution of what open source really means and I am tired of it.
The model is a static data file like a word doc.
Meta open sourced the code to run inference on the model (ie the code for Microsoft word reading a doc file).
They also open sourced the code to train/fine tune the model. (Ie the code for Microsoft word writing a doc file)
Then they released a special doc (the llama 3 model), but didn’t include the script they used to create that doc.
Anything else and you're just open-source-washing your proprietary technology.
Llama did not license its training data. It’s almost impossible to prove a particular LLM was used to generate any particular text, and there’s likely a bunch of illegal content within the dataset used to train the model (as is the case for most other LLMs)…
So why should I care about following a license? They have no mechanism to enforce it. They have no mechanism to detect when it’s being violated. They themselves indicated hostilities to other licenses, so why not ignore it?
i.e. high temperature, exotic samplers (i.e. typicality sampling), using lora/soft prompts/representation engineering, using it as part of a chain on top of other AI, etc
I don't care if every lawyer on earth is hired by Meta. Show me evidence that any particular LLM can be trivially fingerprinted based on its outputs. No, that "red token green token" paper on watermarking (https://arxiv.org/pdf/2301.10226) is not an example of this because of how trivial it is to defeat.
Edit: I can't reply to the comment saying "Subpoena" but this commentator seems to think that using any LLM at all is grounds for a court to issue a Subpoena requiring you to disclose which LLM you're using. If this actually happened, you'd see a massive chilling effect. Also, what stops someone from silently replacing the model with a non infringing one the moment someone starts asking questions?
I'm pretty sure that most courts aren't capable of getting expert testimony which is good enough to deduce that I silently swapped out my blarg_3.1 model which was made using a 1/3 llama3 merge with something else with gloop_1.5 which is no longer infringing.
Like seriously, I again ask, given the idea of courts with warrants and Subpoena's, why should I care about meta's licensing?
Edit2: If you're afraid of an employee leaking this info, don't tell your employees. Good thing clever model merging leaves no traces if you delete metadata!
Subpoenas.
https://www.forbes.com/sites/alexandralevine/2023/12/20/stab...
https://www.airforcetimes.com/news/your-air-force/2024/03/26... (i.e. as source of where LLMs get their classified data from)
If you scrape a large enough part of the internet, you're naturally going to get extremely illegal training data that you won't effectively filter out. I guarantee you that at least a tiny bit of highly classified information was not filtered out of most LLM training data (it wasn't found during the filter step), and it's quite remarkable that this and the above revelations have not led to anyone being Subpoena'd or related in regards to it.
So no, I think that folks will be literally the opposite of litigious on this issue. You want to play that game Zuck? Let's see what happens when I hire my researchers to find the dirt on your models dataset. We will see then who "settles out of court".
Meanwhile, all those lawyers need is to grab your text messages, emails, and slack messages. There's no need to look at code. If you have been up to something it's likely to come out pretty quickly.
p.s., Hiding your tracks after a subpoena is issued is a good way to end up in jail, at least in the US.
https://en.wikipedia.org/wiki/Trap_street
We don't have a commonly-accepted definition of what open source means for LLMs. Just negating Facebook's doesn't advance the discussion.
The open-source community is fractured between those who want it to mean weights available (Facebook); weights and transformer available; weights with no use restrictions (I think this is you); and weights, transformer and training data with no restrictions (obviously not workable, not even the OSI's proposed definition goes that far [1]).
In a world where the only LLMs are proprietary or Llama and the open-source community either remains fractured or chooses an unworkable ask, the latter wil define how the term is used.
[1] https://opensource.org/deepdive/drafts/open-source-ai-defini...
If you distribute the output of bison (say, foo.c) and not the foo.y sources, you would get pushback.
Then there is the EULA which makes it closed source right from the start.
This makes an LLM (emphasis on large) that is open source per this definition legally impossible. In every jurisdiction of consequence.
People like the term open source. It will get used. It is currently undefined. If the choice is an impractical definition and a bad one, we'll get stuck with the bad one. (See: hacker v cracker, crypto(currency) v crypto(graphy), et cetera.)
Maybe. But this isn't how language works. Particularly not English.
A corollary of No true Scotsman [1] is the person administering that purity test rarely gets to define a Scotsman.
[1] https://en.wikipedia.org/wiki/No_true_Scotsman
When you see a MIT-licensed repository, do you call a screenshot, icons or other image assets "open pictures"? That would be laughable. We don't need non-ambiguity outside of context.
Second, the purism argument is just silly. Is GCC not open source because there's no LaTeX for IA-32 manual, and its x86 machine instruction generator is clearly distilled from it?
Even your reply to this one lacks nuance that I brought up. Where do you draw a line for the things that need "source" vs the ones that don't?
It has nothing to do with natural language evolution.
LLMs were entirely proprietary. In that context, Facebook put forward a foundational model one can run on their own hardware and called it open.
At the time, nobody had defined what an open-source LLM was. People came out to say Llama wasn't open. But nobody rigorously proposed a practical definition. The OSI has started, but they're being honest about it being a draft. In the meantime, people are organically discussing open versus proprietary models in a variety of contexts, most of which aren't particularly concerned with the OSI's definitions.
one is pragmatic, ergonomic, and motivated by advancing relationships between communicating persons in a naturally occurring way because the common denominator is a quick race to mutual understanding.
the other is manufactured, and not motivated by relation and advancing communication, but by how the shift in understanding benefits the one pushing for it, and often involves telling people how to think but is doing it through subversion.
Sort of. I'm claiming the meaning of open source when applied to AI is unsettled. There are guiding principles that seem to imply Llama is not open source. But merely pointing that out without offering a practical alternative definition almost guarantees that the intentionally-misused definition Facebook is promulgating becomes the accepted one.
however i do not think the alternatives need to be proposed at this moment in time because right now, the discussion is about holding people who intentionally reframe and misuse words accountable for their "double-speak" given the term's precedent.
conventionally, and by historical collective understanding, it is not open source.
i get you are attempting to highlight AI perhaps means this should be a definition reconsidered, but the irony here is the message itself is distorted due to the conflation, hence why consistency in language to me seems self-evident as a net good.
there is most certainly a difference between naturally occurring (which our brains reeeally support in terms of language development and symbolic communication), and manufactured (and therefore pushing for a word, or more aptly, a perspective's adoption).
i'd rather words manifest through a common need to reach mutual understanding as a means to relate to one another and this world, rather than having someone who stands to benefit from the change in definition, tell me what it means, and then expect me to just "agree", while they campaign around that and pretend it's the established definition (and not actually their own revised version).
it'd be one thing if people who were throwing the term around so loosely would be transparent: "Hey, we know this isn't historically what everyone means by OSS, but... that's OSS your OSS this is OSS, everything is OSS
instead a lot of these narratives are standing on the shoulders of the original definition and context of what it means to be OSS, and therefore the pedigree, implications (and whatever else for PR spin/influence), and simultaneously diluting what it means in the process as the definition gets further and further obfuscated by those influencing the change, and its pedigree is relied on as a distraction away from what is being done, or actually said.
They could use the more correct Open Weights (which is still a euphemism because of the EULA).
But they do not, and they know perfectly well what they are doing. They are the ones responsible for these discussions, but they double down and blame the true OSS people.
Of course it is. Look at this thread. Look at the policy discussions around regulating AI. Hell, look at the OSI's draft definition [1].
Pretending something is rigorously defined the way you want it to be doesn't make it so.
[1] https://opensource.org/deepdive/drafts/open-source-ai-defini...
"Open-source" represents a cluster of concepts but at the core of it there is a specific definition in spirit at least -- you can see the source for yourself, and compile it for yourself.
If the source is not available, why would you want to call it open-source? Just call it something else. As simple as that.
Definitions flow out of usage. The definition clarifies how the word is used and what people mean when they do use it.
You are, in a very literal sense, doing what Orwell, et al was so desperately against by actively controlling how language is permitted to be used.
Nobody can "force" someone else to use the correct definitions of words, but when people disregard their established meanings they risk communication breaking down and the confusion and misunderstandings that follow. If I went around speaking nonsense or making up my own invented definitions for established words I shouldn't expect to be understood and others would be perfectly right to correct me or ask that I stick to using the well understood and documented meaning of words if I expect to have a productive conversation.
It's also perfectly fair to call out people who twist the meaning of words intentionally so that they can lie, mislead, and manipulate others. When it comes to products, companies can't just say "Words can mean anything I say they do! There are no rules!" to get away with false advertising.
The meaning of words drifts in every living language.
> perfectly fair to call out people who twist the meaning of words intentionally
We don't have consensus around what open source means for LLMs. Facebook is pretending we do. But so is everyone in this thread claiming there is a single true definition of an open source LLM.
And it does result in a lot of confusion and misunderstanding until gradually people are taught the new definitions and how they are used. There are also groups of people who deliberately and continuously redefine words because they don't want to be widely understood. Some want to develop a means to signal to and identify others within their in-group, and some want to keep outsiders from understanding them so they can speak more openly in mixed company.
> We don't have consensus around what open source means for LLMs.
There are people who will argue about what open source means for anything. It's okay that open source means different things to different people, but it does result in confusion in discussions until people make their definitions clear.
I don't think that Facebook has earned the benefit of the doubt, in fact they've more than earned our skepticism, so it's very reasonable to see their new definition of "open source" as being nothing but marketing rhetoric at best, or at worst, as an attempt to twist our still developing consensus on what open source means into something that violates the philosophy/spirit of the open source movement.
Re-definition example: War is peace, freedom is slavery.
Since the new euphemism for downloadable models is a re-definition, Orwell would have been 100% against it. In fact the new use of "open source" is an Orwellian term.
i'm inclined to believe Orwell would have disagreed with OP, and would be asking himself -- why is there such a distinct push by those who benefit from the reframing, to reframe what Open Source means (compared to its already established meaning).
Orwell highlighted and warned against the consequences of people in influential positions of power intentionally distorting the collective associations with their new, updated versions of existing words, campaigning around those distortions, and intentionally reframing associations over time such that, the associations are polarizing and obfuscated, motivated by manipulation to benefit a select few, not motivated by advancing communication -- it certainly was not an example of society and its language naturally evolving "definitions" through usage.
and the novel wasn't a criticism against slang, or association/vernacular changing/evolving over time throughout collective use, nor was it a stance on requiring fixed, permanent, unwavering definitions -- it only emphasized how important it is to have consistent meaning.
he just wanted to encourage people to be skeptical of those pushing for the "different" or updated meaning of words, that clearly had a well-defined context, and association, previously -- why are they so dedicated and determined to "push" for a new meaning to get accepted, when there is a previously established and well accepted meaning already.
that doesn't sound natural to me, that sounds manufactured.
GPLv3 defines "source code" as the preferred form for making changes.
For most normal software that is identical to what you'd use to recreate it... but the way to make changes to an LLM isn't to rebuild it, but is to run fine-tuning on it.
Once again: for those who are new here. There is Free Software, which has a usefully strict definition.
And there is Open Source, the business-friendly -- but consequently looser -- other thing.
You can like one or both, they both have advantages and drawbacks.
But you cannot insist that "Open Source" has a very strict definition. It just doesn't. That's why the whole Free Software thing is needed, and IMHO, more important.
Statements like this remind me that I really need to KEEP GOING with this.
One could argue that the OSI should have gotten a trademark on the term. But the FSF doesn't have a trademark on the term "free Software" either, so the terms have approximately equal legal protections.
Meta using the term "open source" to apply to their model data when their license isn't an open source license is dishonest at best.
Free Software has the GPL, and all its related, healthy controversy. It's not perfectly clear, but it's far more battle-tested than the much more nebulous "Open source."
People who like "free software" put in work, and better understood that, to some extent, you can't have your cake and eat it too. "Open Source" is much more about a whole lot (to me, naive) wishful thinking.
(The OSI is a bunch of companies in a trenchcoat, the creators of the GPL were more principled.)
I just noticed they are currently discussing what "Open Source AI" should mean. You can join in and add your thoughts tot he discussion.
What? There are so many open source projects from huge companies these days.
VSCode, .NET, typescript from MS
Angular, flutter, kubernetes, go, android, chromium from Google
Yes it is still better than not being able to access the weights at all, but calling these weights open source is not correct.
[0] https://huggingface.co/meta-llama/Meta-Llama-3.1-8B/blob/mai...
But many legal minds close to this issue have moved to the position that there is no meaningful distinction, at least when it comes to licenses like GPL.
For example: https://writing.kemitchell.com/2023/10/13/Wrong-About-GPLs
I don't think Facebook/Meta is the beacon of open-source goodness you think it is. The main reason they created yarn instead of iterating on npm is to use their own patent-friendly license they wanted to use with React (before the community flipped out and demanded they re-license it as MIT). Early Vue adoption seemed mostly driven by that React licensing fiasco.
You can benefit a lot from it, and I have... but do be sure you know what you are ferrying on your back before you decide to offer it a ride across the river.
Also it's bad when HN is downvoting fking GWERN
Not only is "a lot" of FOSS not released like this, both free software and Meta's models cannot be monetized post-release. If Meta decides to charge money for Llama4, then everyone with access to the prior models can keep their access and even finetune/redistribute their model. There is no strategic flip Meta can attempt here without shotgunning their own foot off.
For two - they can't do shit. Every one of those businesses can try to pivot into being the next OpenAI, but even OpenAI can't be fucked to turn a profit. They can't retract their models (remember Llama 1.0? hahahahahahahahaha) and they can't piss and moan to the authorities when their weights leak because 99% of them contain unlicensed copyrighted material in the first place.
The real twist to this Gordian knot? Open source is the only model you can release conventional AI under, because anything else precludes a Fair Use defense. Everyone that has "proprietary" models only manages it by playing keep-away with the community.
> Only a five-year-old child is either surprised or angered by this turn of events.
Five year old children know that they can keep their toys once they stop selling them at Wal-mart. Pulling a model off HF doesn't stop new finetunes or even redistribution and re-quantization of the old model. Mistral and Runway can do whatever they want, all their models are belong to us.
Zero privacy?
I want it to mean something!
They are synonym and mean the same thing.
Also open source means the license used should be something standard not proprietary, without restrictions on how you can use it.
Not at all. They're only similar in the sense that both are a build artifact.
> Without visibility into the training data, curation / moderation decision, the training code, etc Llama could be doing anything and we wouldn’t know.
"could be doing anything" is quite the tortured phrase, there. For one, model training is not deterministic and having the full training data would not yield a byte-perfect Llama retrain. For two, the released models are not turing-complete or filled with viruses; you can open the weights yourself and confirm they're static and harmless. For three, training code exists for all 3 Llama models and the reason nobody uses them is because it's prohibitively expensive to reproduce and has zero positive potential compared to finetuning what we have already.
> Also open source means the license used should be something standard not proprietary, without restrictions on how you can use it.
There are very much restrictions on redistribution for nearly every single Open Source license. Permissive licensing may not mean what you think it means.
My company is building an application for an university client, regarding the examination of research data written in "human language" (mostly notes and docs).
Due the high confidentiality of the subjects, as often they deal with non-patented information, we couldn't risk using frontier models, as it could break the novelty of the invention, therefore losing patentability.
Now with Llama3.1, we can simply run these models locally, on systems that is not even connected to the internet. LLMs are mostly good in examining massive amount of research papers and information, at least for the application we are aiming at, saving thousands of hours of tiresome (and very boring) human labour.
I am trying to endorse Meta or Zuckerberg or anything like that, but at least in this aspect, I think Llama being "open-source" is a very good aspect.
Is the concern that prompts could be re-used for training by the provider and such knowledge become part of the model?
Open-Knowledge source with an Open-Inteligence that can guide you through the entire massive digital library of its own brain. Semantic data light-years beyond a Search Engine.
Kagi’s LLM can already do that. I believe so can Perplexity’s. Citing sources isn’t something only open models can do.
Sure. The point is the thing you said only an open-source model can do, it can do. Plenty of proprietary LLMs can cite sources.
The plain truth is most of the benefits of open models are not on the consumer side. (Or at least, I haven't seen any articulated.) They're on the producers'. Open models are better for those of us training models. That's partly why the open data debate is academic--very few people are training large foundation models because the compute and electricity costs are prohibitive.
Genuinely blown away the EU isn't doing this.
In the U.S., the solution may be in carving a legal safe harbor for companies that release their models per the OSI's draft definition of open source.
This sounds like the usual AI marketing with the word "open" thrown in. It's not articulating something youc an only do with an open source LLM (and doesn't define what that means).
I'm personally not thrilled with how locked down LLMs are. But we'll need to do a better job at articulating (a) a definition and (b) the benefits of adhering to it versus the "you can run it on your own metal" definition Facebook is promulgating. Because a model meeting Facebook's definition has obvious benefits over proprietary models run on someone else's servers.
I believe our world fights to destroy ideas like this because our economy drives our entire life.
>> No
>>> You can't imagine it
You haven’t articulated the idea you claim the “world fights to destroy”. (Just throwing around the word open without elaboration isn’t an idea.)
No. But you haven’t articulated why making everyone’s Facebook chats public is a net good. What does opening that data up confer in practical benefits?
Given what we know about LLMs, one trained only on public-domain data will underperform one trained on that plus proprietary data. If you want source data available, you have to either concede the "open" models will be structurally handicapped or that all data must be public.
Facebook says no, at least for Llama [1].
[1] https://itlogs.com/facebook-uses-user-data-to-train-ai-but-l...
We need dramatic reforms on copyright, as we’ve really let corporate interests crowd out our rights to human culture and ideas. While I alone cannot decide what we as a country should find reasonable, I can say I find 20 years + 5 years extension is perfectly reasonable and that corporations should have never been able to pay off politicians to get what they wanted. Let alone Sonny Bono, that bastard, signing in bills that specifically benefited him.
So, to reiterate, the idea I feel that corporations want to destroy is the idea that we, as a people, have rights to the works that form our popular culture and that no one man, let alone a faceless corporation, should be able to profit from a singular work for hundreds of years.
Also just to be clear, if you want to set up a RAG with an open weight model and a large dataset there's nothing stopping you. Download Red Pajama and Llama and give it a try.
https://github.com/togethercomputer/RedPajama-Data
OTOH why you want the data is not clear. You don't need it to run Meta'a models for free, or to fine-tune them for your own needs. The only thing the data would allow you is to pre-train from scratch, in other words to obtain the exact same set of weights that Meta is giving you for free.
They are never going to stop saying this or show us the actual source data. Imagine if they did... Do they even entertain the idea? Can they really not imagine Open Source AI being possible because of all the personal data they train on?
Looking forward to continued updates and releases of Llama (and SAM!) from Meta.
[1] https://www.theverge.com/2024/8/28/24229068/california-sb-10...
It is disheartening so see the damage capacity in the hands of a couple of paranoic people who perhaps read the wrong scifi and had lots of power to influence others. If California passes this law, in a few years the world economy will be very different.
Sure they will. Just like how the auto manufacturers stopped trying to sell in CA after stricter emissions regulations and just like how companies stopped trying to serve European visitors after GDPR came into force. California is a $4T GDP, placing it just behind Japan and ahead of India in global rankings. I am sure they are happy to tell non-complaint companies to piss off and not let the door hit them on the way out. This is the canary in the coal mine for cowboy AI companies, once EU regulators get into the game their days of playing fast and loose with this tech will be over.
Here are some counter-points:
Regulation:
- Very little effort is made to evaluate risk of over-regulating, regulatory capture and counterproductive wrong regulation
- The downside of under-regulating is vastly overemphasized, most arguments boil down to "we have to act FAST now or x BAD thing might happen"
- The risk of over/wrongly regulating is vastly under emphasized with the same FUD reasons.
- according to one of the many straw-men argument in the pod I'm a libertarian against any and all regulation because I criticize possible regulatory capture, I would enthusiastically support regulation that foundation models have to be:
-- given freely to public researchers/academics for in-depth independent safety-research
-- open weighted after a while (e.g. after ~ a year, safety concerns should be mostly ruled out and new generations are out so ROI is already likely there. [e.g. there's NO safety reason at all for ClosedAI to not release gpt-3.5, llama3 is better already])
Proposed FLOP cut-off of SB 1047:
- according to the pod, the cut off is much more advanced than anything currently released.
- The 10^26 FLOP cutoff is way to low, llama-405b is ~4×10^25 FLOPs
- 405B is maybe 20% smarter than 70B, while taking over an order of magnitude more FLOPS to train, the cutoff itself is very likely not much smarter than the current SOTA.
- IMO none of the current SOTA models are very dangerous, but kill switch regulation is.
Kill-Switches:
- SB 1047 is (non-explicitly) calling for kill-switches over the cut-off due to liability of the model creators and market dynamics
- Any kill-switch regulation means a complete dead-end to any advanced open-weights AI. This means that huge corporations and governments will control any and all advanced AI-development. This is top-down control of the maths you are allowed to run on your computer IMO that is Orwellian as fuck.
China:
- mentioning china is FUD 101, it's basically AI's "think of the children"
- If they think they can stop china from building their own advanced LLMs, they're delusional. This regulation might even help them to get there faster. They don't even need to steal, there maybe a year or two behind the SOTA and catching up fast.
I just don't get how so many people on a site with "hacker" in the name want to make it impossible to hack on these things for anyone not employed by the big corporate AI research labs.
Disturbing approach to mitigating AI harms imo, this bill basically hopes it can limit the number of operators of an arbitrary model type so as to allow easier governance of AI model use. This ignores the reality that we already have large models openly released and easily modifiable (outside CA jurisdiction) which likely are capable of perpetuating “critical harms”, or that the information requirements to achieve the defined “critical harms” could be realized by an individual by simply reading a few books. There’s also no reason to simply assume that future models will require millions of dollars of compute to create; fulfilling the goals of this regulatory philosophy long term almost certainly requires the banning of general purpose compute to come close to the desired outcome of a supposed reduction in probability of some “critical harm” being perpetrated. We should be focusing on hardening society to the realities of the “critical harms” identified by this bill, rather than implicitly assuming the only reason we don’t see them as much irl is because everyone is stupid. The current paranoia wave around LLMs is just a symptom of people waking up to the fragility of the world we live in.