You also can&#x27;t use it for training or improving other models.You also can&#x27;t use it if you&#x27;re the government of India.Neither can sex workers use it. (Do you know if your customers are sex workers?)There are also very vague restrictions for things like discrimination, racism etc.

&quot;Use it for whatever you want(conditions apply), but not if you are Google, Amazon, etc. If you become big enough talk to us.&quot; That&#x27;s how I read the license, but obviously I might be missing some nuance.

Also it isn&#x27;t source code, it is a binary. You need at least the data curation code and preferably the data itself for it to be actually source code in the practical sense that anyone can remake the build.Llama could change the license on later versions to kill your business and you have no options as you don&#x27;t know how they trained it or have the budget to.It&#x27;s not much more free than binary software.

Sure but under what license?
Because slapping “open source” on the model doesn’t make it open source if it’s not actually license that way. The 3.1 license still contains their non-commercial clause (over 700m users) and requires derivatives, whether fine tunings or trained on generated data, to use the llama name.

And this is happening RIGHT as a new potential leader is emerging in Llama 3.1. I&#x27;m really curious about how this is going to match up on the leaderboards...

Interesting quotes. Less sophisticated actors just means humans who already write in 2020 what the NYT wrote in early 2022 to prepare for Biden&#x27;s State Of The Union 180° policy reversals (manufacturing consent).FB was notorious for censorship. Anyway, what is with the &quot;actions&#x2F;actors&quot; terminology? This is straightforward totalitarian language.

&quot;AI detection doesn&#x27;t work, and with how much better the systems are getting, it&#x27;s probably never going to, unless you keep the best models for yourself&quot;I don&#x27;t think that&#x27;s true. I don&#x27;t think even the best privately held models will be able to detect AI text reliably enough for that to be worthwhile.

I found this dubious as well, especially how it is portrayed as a simple game of compute power. For a start, there is an enormous asymmetry which is why we have a spam problem in the first place. For example a single bot can send out millions of emails at almost no cost and we have to expend a lot more &quot;energy&quot; to classify each one and decide if it&#x27;s spam or not. So you don&#x27;t just need more compute power you need drastically more compute power, and as AI models improve and get refined, the operation at ten times the scale is probably going to be marginally better, not orders of magnitude better.I still agree with his general take - bad actors will get these models or make them themselves, you can&#x27;t stop it. But the logic about compute power is odd.

&gt; This is how we’ve managed security on our social networks – our more robust AI systems identify and stop threats from less sophisticated actors who often use smaller scale AI systems.Ok, first of all, has this really worked? AI moderators still can&#x27;t capture the mass of obvious spam&#x2F;bots on all their platforms, threads included. Second, AI detection doesn&#x27;t work, and with how much better the systems are getting, it&#x27;s probably never going to, unless you keep the best models for yourself, and it&#x27;s is clear from the rest of the note that its not zuck&#x27;s intention to do so.&gt; As long as everyone has access to similar generations of models – which open source promotes – then governments and institutions with more compute resources will be able to check bad actors with less compute.This just doesn&#x27;t make sense. How are you going to prevent AI spam, AI deepfakes from causing harm with more compute? What are you gonna do with more compute about nonconsensual deepfakes? People are already using AI to bypass identity verification on your social media networks, and pump out loads of spam.

I think you are asking what the benefits are? The main benefit is that we can trust what these systems are doing better. Or we can self host them. If we just take the weights, then it is unclear how these systems might be lying to us or manipulating us.Another benefit is that we can learn from how the training and other steps actually work. We can change them to suit our needs (although costs are impractical today). Etc. It’s all the usual open source benefits.

&gt; Only if it is truly open source (open data sets, transparent curation&#x2F;moderation&#x2F;censorship of data sets, open training source code, open evaluation suites, and an OSI approved open source license)You’re missing a then to your if. What happens if it’s “truly” open per your definition versus not?

There is also the risk of companies like Meta introducing ads in the training itself, instead of inference time.

Yeah, though I do wonder for a big model like 405B if the original training recipe, really matters for where models are heading, practically speaking which is smaller and more specific?I imagine its main use would be to train other models by distilling them down with LoRA&#x2F;Quantization etc(assuming we have a tokenizer). Or use them to generate training data for smaller models directly.But, I do think there is always a way to share without disclosing too many specifics, like this[1] lecture from this year&#x27;s spring course at Stanford. You can always say, for example:- The most common technique for filtering was using voting LLMs (without disclosing said llms or quantity of data).- We built on top of a filtering technique for removing poor code using ____ by ____ authors (without disclosing or handwaving how you exactly filtered, but saying that you had to filter).- We mixed certain proportion of this data with that data to make it better (without saying what proportion)[1] <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=jm2hyJLFfN8&amp;list=PLoROMvodv4rNiJRchCzutFw5ItR_Z27CM&amp;index=32" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=jm2hyJLFfN8&amp;list=PLoROMvodv4...</a>

Only if it is truly open source (open data sets, transparent curation&#x2F;moderation&#x2F;censorship of data sets, open training source code, open evaluation suites, and an OSI approved open source license).Open weights (and open inference code) is NOT open source, but just some weak open washing marketing.The model that comes closest to being TRULY open is AI2’s OLMo. See their blog post on their approach:<a href="https:&#x2F;&#x2F;blog.allenai.org&#x2F;hello-olmo-a-truly-open-llm-43f7e7359222" rel="nofollow">https:&#x2F;&#x2F;blog.allenai.org&#x2F;hello-olmo-a-truly-open-llm-43f7e73...</a>I think the only thing they’re not open about is how they’ve curated&#x2F;censored their “Dolma” training data set, as I don’t think they explicitly share each decision made or the original uncensored dataset:<a href="https:&#x2F;&#x2F;blog.allenai.org&#x2F;dolma-3-trillion-tokens-open-llm-corpus-9a0ff4b8da64" rel="nofollow">https:&#x2F;&#x2F;blog.allenai.org&#x2F;dolma-3-trillion-tokens-open-llm-co...</a>By the way, OSI is working on defining open source for AI. They post weekly updates to their blog. Example:<a href="https:&#x2F;&#x2F;opensource.org&#x2F;blog&#x2F;open-source-ai-definition-weekly-update-july-15" rel="nofollow">https:&#x2F;&#x2F;opensource.org&#x2F;blog&#x2F;open-source-ai-definition-weekly...</a>

&gt; Doubtful that GPUs purchased today would be in use for a similar time scaleTotally agree. That doesn&#x27;t mean it can&#x27;t generate massive ROI.&gt; Govt investment would also drive the cost of GPUs up a great dealDifficult to say this ex ante. On its own, yes. But it would displace some demand. And it could help boost chip production in the long run.&gt; Not sure why a publicly accessible GPU cluster would be a better solution than the current system of research grantsThose receiving the grants have to pay a private owner of the GPUs. That gatekeeping might be both problematic, if there is a conflict of interests, and inefficient. (Consider why the government runs its own supercomputers versus contracting everything to Oracle and IBM.)

Of course they won&#x27;t. The investment in the Heavy Press Program was the initial build, and just citing one example, the Alcoa 50,000 ton forging press was built in 1955, operated until 2008, and needed ~$100M to get it operational again in 2012.The investment was made to build the press, which created significant jobs and capital investment. The press, and others like it, were subsequently operated by and then sold to a private operator, which in turn enabled the massive expansion of both military manufacturing, and commercial aviation and other manufacturing.The Heavy Press Program was a strategic investment that paid dividends by both advancing the state of the art in manufacturing at the time it was built, and improving manufacturing capacity.A GPU cluster might not be the correct investment, but a strategic investment in increasing, for example, the availability of training data, or interoperability of tools, or ease of use for building, training, and distributing models would probably pay big dividends.

A much better investment would be to (somehow) revolutionize production of chips for AI so that it&#x27;s all cheaper, more reliable, and faster to stand up new generations of software and hardware codesign. This is probably much closer to the program mentioned in the top level comment: It wasn&#x27;t to produce one type of thing, but to allow better production of any large thing from lighter alloys.

&gt; Not sure why a publicly accessible GPU cluster would be a better solution than the current system of research grants.You mean a better solution than different teams paying AWS over and over, potentially spending 10x on rent rather than using all that cash as a down payment on actually owning hardware? I can&#x27;t really speak for the total costs of depreciation&#x2F;hardware maintenance but renting forever isn&#x27;t usually a great alternative to buying.

Doubtful that GPUs purchased today would be in use for a similar time scale. Govt investment would also drive the cost of GPUs up a great deal.Not sure why a publicly accessible GPU cluster would be a better solution than the current system of research grants.

&gt; any public cluster would be outdated in 2 yearsThe private companies buying hundreds of billions of dollars of GPUs aren&#x27;t writing them off in 2 years. They won&#x27;t be cutting edge for long. But that&#x27;s not the point--they&#x27;ll still be available.&gt; Nvidia&#x27;s profit margins on the H100 are crazyI don&#x27;t see how the current practice of giving a researcher a grant so they can rent time on a Google cluster that runs H100s is more efficient. It&#x27;s just a question of capex or opex. As a state, the U.S. has a structual advantage in the former.&gt; far better ways to spend $1b in research dollarsOne assumes the U.S. government wouldn&#x27;t be paying list price. In any case, the purpose isn&#x27;t purely research ROI. Like the heavy presses, it&#x27;s in making a prohibitively-expensive capital asset generally available.

What about dollar cost averaging your purchases of GPUs? So that you&#x27;re always buying a bit of the newest stuff every year rather than just a single fixed investment in hardware that will become outdated? Say 100 million a year every year for 20 years instead of 2 billion in a single year?

The problem is that any public cluster would be outdated in 2 years. At the same time, GPUs are massively overpriced. Nvidia&#x27;s profit margins on the H100 are crazy.Until we get cheaper cards that stand the test of time, building a public cluster is just a waste of money. There are far better ways to spend $1b in research dollars.

Not--to my knowledge--for the GPUs necessary to train cutting-edge LLMs.

This is the most recent iteration of a national platform. They have tons of GPUs (and CPUs, and flash storage) hooked up as a Kubernetes cluster, available for teaching and research.<a href="https:&#x2F;&#x2F;nationalresearchplatform.org&#x2F;" rel="nofollow">https:&#x2F;&#x2F;nationalresearchplatform.org&#x2F;</a>

Don&#x27;t these public clusters exist today, and have been around for decades at this point, with varying architectures? In the sense that you submit a proposal, it gets approved, and then you get access for your research?

&gt; NSF has had a bunch of different programs to allocate time on supercomputers to researchers at no cost, these days mostly run out of the Office of Advanced CyberinfrastrutureThis would be the logical place to put such a programme.

As you&#x27;ve rightly pointed out, we have the mechanism, now let&#x27;s fund it properly!I&#x27;m in Canada, and our science funding has likewise fallen year after year as a proportion of our GDP. I&#x27;m still benefiting from A100 clusters funded by tax payer dollars, but think of the advantage we&#x27;d have over industry if we didn&#x27;t have to fight over resources.

Yeah, the specific AI&#x2F;ML-focused program is NAIRR.<a href="https:&#x2F;&#x2F;nairrpilot.org&#x2F;" rel="nofollow">https:&#x2F;&#x2F;nairrpilot.org&#x2F;</a>Terrible name unless they low-key plan to make AI researchers&#x27; hair fall out.

the US already pays for 2+ aws region for cia&#x2F;dod. why not pay for a region that is only available to researchers?

The National Science Foundation has been doing this for decades, starting with the supercomputing centers in the 80s. Long before anyone talked about cloud credits, NSF has had a bunch of different programs to allocate time on supercomputers to researchers at no cost, these days mostly run out of the Office of Advanced Cyberinfrastruture. (The office name is from the early 00s) - <a href="https:&#x2F;&#x2F;new.nsf.gov&#x2F;cise&#x2F;oac" rel="nofollow">https:&#x2F;&#x2F;new.nsf.gov&#x2F;cise&#x2F;oac</a>(To connect universities to the different supercomputing centers, the NSF funded the NSFnet network in the 80s, which was basically the backbone of the Internet in the 80s and early 90s. The supercomputing funding has really, really paid off for the USA)

It is an exciting concept, there&#x27;s a huge wealth of gaming hardware deployed that is inactive at most hours of the day. And I&#x27;m sure people are willing to pay well above the electricity cost for it.Unfortunately, the dominant LLM architecture makes it relatively infeasible right now.- Gaming hardware has too limited VRAM for training any kind of near-state-of-the-art model. Nvidia is being annoyingly smart about this to sell enterprise GPUs at exorbitant markups.- Right now communication between machines seems to be the bottleneck, and this is way worse with limited VRAM. Even with data-centre-grade interconnect (mostly Infiniband, which is also Nvidia, smart-asses), any failed links tend to cause big delays in training.Nevertheless, it is a good direction to push towards, and the government could indeed help, but it will take time. We need both a more healthy competitive landscape in hardware, and research towards model architectures that are easy to train in a distributed manner (this was also the key to the success of Transformers, but we need to go further).

Ever heard of SETI@home?<a href="https:&#x2F;&#x2F;setiathome.berkeley.edu" rel="nofollow">https:&#x2F;&#x2F;setiathome.berkeley.edu</a>

What about distributed training on volunteer hardware? Is that feasible?

It would be probably cheaper to negate some IP. There are quite some projects and initiatives to make CUDA code run on AMD for example, but as far as I know, they all stopped at some point, probably because of fear of being sued into oblivion.

It seems like rocm is already fully ready for transformer inference, so you are just referring to training?

Please start with the Windows Tax first for Linux users buying hardware...and the Apple Tax for Android users...

Either you port Tensorflow (Apple)[1] or PyTorch to your platform or you allow CUDA to run on your hardware (AMD) [2]. Companies are incentives to not have NVIDIA having a monopoly but the thing is that CUDA is a huge moat due to compatibility of all frameworks and everyone knows it. Also, all of the cloud or on premises providers use NVIDIA regardless.[1] <a href="https:&#x2F;&#x2F;developer.apple.com&#x2F;metal&#x2F;tensorflow-plugin&#x2F;" rel="nofollow">https:&#x2F;&#x2F;developer.apple.com&#x2F;metal&#x2F;tensorflow-plugin&#x2F;</a>
[2] <a href="https:&#x2F;&#x2F;www.xda-developers.com&#x2F;nvidia-cuda-amd-zluda&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.xda-developers.com&#x2F;nvidia-cuda-amd-zluda&#x2F;</a>

That&#x27;s the kind of work that can come out of academia and open source communities when societies provide the resources required.

It is being done already...<a href="https:&#x2F;&#x2F;docs.scale-lang.com&#x2F;" rel="nofollow">https:&#x2F;&#x2F;docs.scale-lang.com&#x2F;</a>

Or just develop the next wave of chips designed for specifically transformer-based architectures (and ternary computing), and bypass the needs for GPUs and CUDA altogether

How about using some of that money to develop CUDA alternatives so everyone is not paying the Nvidia tax?

Great idea, too bad the DOE and NSF were there first.

The size of the cluster would have to be massive or else your job will be on the queue for a year. And even then what are you going to do downsize the resources requested so you can get in earlier? After a certain point it starts to make more sense to just buy your own xeons and run your own cluster.

Imagine if they made a data center with 1957 electronics that cost $279 million.They probably won&#x27;t be using it now because the phone in your pocket is likely more powerful. Moore law did end but data center stuff are still evolving order of magnitudes faster than forging presses.

Yeah this would be the real equivalent of the program people are talking about above. That an investing in core networking infrastructure (like cables) instead of just giving huge handouts to certain corporations that then pocket the money.....

I&#x27;d like to see big programs to increase the amount of cheap, clean energy we have. AI compute would be one of many beneficiaries of super cheap energy, especially since you wouldn&#x27;t need to chase newer, more efficient hardware just to keep costs down.

&gt; If anything, allocate compute to citizens.If something like this were to become a reality, I could see something like &quot;CitizenCloud&quot; where once you prove that you are a US Citizen (or green card holder or some other requirement), you can then be allocated a number of credits every month for running workloads on the &quot;CitizenCloud&quot;. Everyone would get a baseline amount, from there if you can prove you are a researcher or own a business related to AI then you can get more credits.

So we&#x27;ll have the government bypass markets and force the working class to buy toys for the owning class?If anything, allocate compute to citizens.

The American healthcare industry demonstrates the tremendous benefits of rigidly applying this mindset.Why couldn’t law enforcement be private too? You call 911, several private security squads rush to solve your immediate crime issue, and the ones who manage to shoot the suspect send you a $20k bill. Seems efficient. If you don’t like the size of the bill, you can always get private crime insurance.

That’s not correct. The American health care system is an extreme example of where private organisations fail overall society.

&gt; Overall government doing anything is a bad idea.that is bereft of detail enough to just be wrong. There are things that government is good for and things that government is bad for, but &quot;anything&quot; is just too broad, and reveals an anti-government bias which just isn&#x27;t well thought out.

Why are governments a bad idea? Seems the human race has opted for governments doing things since the dawn of civilization. Building roads, providing defense, enforcing rights, provide social safety nets, funding costly scientific endeavors.

Ugh.Government distorting undeveloped markets that have a lot of room for competition to increase efficiencies is a bad thing.Government agencies running programs that should not be profitable, or where the only profit to be left comes at the expense of society as a whole, is a good thing.Lots of basic medicine is the go to example here, treating cancer isn&#x27;t going to be &quot;profitable&quot; and attempting to make it such just leads to dead people.On the flip side, one can argue that dentistry has seen amazing strides in affordability and technological progress through the free market. From dental xrays to improvements in dental procedures to make them less painful for the patients.Eye surgery is another area where competition has lead to good consumer outcomes.But life of death situations where people can&#x27;t spend time researching? The only profit there comes through exploiting people.

To summarise: There are some things where government action is the best solution, however by default see if the private sector can sort it first.

Overall government doing anything is a bad idea. There are cases however where government is the only entity that can do certain things. These are things that involve military, law enforcement etc. Outside of this we should rely on private industry and for-profit industry as much as possible.

<a href="https:&#x2F;&#x2F;loc.gov&#x2F;pictures&#x2F;search&#x2F;?q=Photograph:%20oh1540&amp;fi=number&amp;op=PHRASE&amp;va=exact&amp;co%20=hh&amp;st=gallery&amp;sg%20=%20true" rel="nofollow">https:&#x2F;&#x2F;loc.gov&#x2F;pictures&#x2F;search&#x2F;?q=Photograph:%20oh1540&amp;fi=n...</a>

I just watched this 1950s DoD video on the heavy press program and highly recommend it: <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=iZ50nZU3oG8" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=iZ50nZU3oG8</a>

Does it? You&#x27;re looking at a gargantuan investment in terms of money that would also require thousands of staff.That just doesn&#x27;t seem a good idea.

It makes much more sense to invest in a next generation fab for GPUs than to buy GPUs and more closely matches this kind of project.

Now, I have no idea.How much capability would $3.2bn in terms of AI computing power provide, including the operational and power costs of the cluster?Certainly, you could build a &quot;$3.2bn GPU cluster&quot;, but it would be dark.So, how much learning time would $3.2bn provide? 1 year? 10 years?Just curious about hand wavy guesses. I have no idea the scope of the these clusters.

Very much in this spirit is the NSF-funded National Deep Inference Fabric, which lets researchers run remote experiments on foundation models: <a href="https:&#x2F;&#x2F;ndif.us" rel="nofollow">https:&#x2F;&#x2F;ndif.us</a>. They just announced a pilot program for Llama405b!

Better idea would be to make various open source packages utilities and put maintainers everywhere funded by public good.AI is a fad, the brick and mortar of the future is open source tools.

For the DoE, take a look at:<a href="https:&#x2F;&#x2F;doeleadershipcomputing.org&#x2F;" rel="nofollow">https:&#x2F;&#x2F;doeleadershipcomputing.org&#x2F;</a>

Snellius is a nice resource. A powerful Slurm based HTC cluster with different cues for different workloads (cpu&#x2F;genomics, gpu&#x2F;deep learning).To access the resource I had to go through EuroCC [0], which is a network facilitating access to and exploitation of HPC&#x2F;HTC infra. It is (or can be) a great competing model to US cloud providers.As a small business I got 8 hrs of consultancy and 10k compute hours for free. I’m still learning the details but my understanding is is that after that the prices are very competitive.[0] <a href="https:&#x2F;&#x2F;www.eurocc-access.eu&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.eurocc-access.eu&#x2F;</a>

Italy built the Leonardo HPC cluster, it&#x27;s one of the largest in EU and was created by a consortium of universities. After just over a year it&#x27;s already at full capacity and expansion plans have been anticipated because of this.

In the Netherlands, for instance, there is &quot;the national supercomputer&quot; Snellius: <a href="https:&#x2F;&#x2F;www.surf.nl&#x2F;en&#x2F;services&#x2F;snellius-the-national-supercomputer" rel="nofollow">https:&#x2F;&#x2F;www.surf.nl&#x2F;en&#x2F;services&#x2F;snellius-the-national-superc...</a>
I am not sure about its budget, but my impression as a user is that its resources are never fully used. At least, none of my jobs ever had to queue.
I doubt that it can compete with the scale of resources that FAANG companies have available, but then again, I also doubt how research would benefit.Sure, academia could build LLMs, and there is at least one large-scale project for that: <a href="https:&#x2F;&#x2F;gpt-nl.com&#x2F;" rel="nofollow">https:&#x2F;&#x2F;gpt-nl.com&#x2F;</a>
On the other hand, this kind of models still need to demonstrate specific scientific value that goes beyond using a chatbot for generating ideas and summarizing documents.So I fully agree that the research budget cuts in the past decades have been catastrophic, and probably have contributed to all the disasters the world is currently facing. But I think that funding prestigious super-projects is not the best way to spend funds.

Eric Schmidt advocated for this exact thing in an Op-ed piece in the latest MIT Technology Review.[1] <a href="https:&#x2F;&#x2F;www.technologyreview.com&#x2F;2024&#x2F;05&#x2F;13&#x2F;1092322&#x2F;why-america-needs-an-apollo-program-for-the-age-of-ai&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.technologyreview.com&#x2F;2024&#x2F;05&#x2F;13&#x2F;1092322&#x2F;why-amer...</a>

&gt; A public cluster of GPUs provided for free to American universities, companies and non-profits might not be a bad idea.USA and Europe is already doing that in a grand scale, in different forms. Both at national and international scale.I work at an HPC center which provides servers nationally and collaborates on international level.

“The Heavy Press Program was a Cold War-era program of the United States Air Force to build the largest forging presses and extrusion presses in the world.” This ”program began in 1944 and concluded in 1957 after construction of four forging presses and six extruders, at an overall cost of $279 million. Six of them are still in operation today, manufacturing structural parts for military and commercial aircraft” [1].$279mm in 1957 dollars is about $3.2bn today [2]. A public cluster of GPUs provided for free to American universities, companies and non-profits might not be a bad idea.[1] <a href="https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Heavy_Press_Program" rel="nofollow">https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Heavy_Press_Program</a>[2] <a href="https:&#x2F;&#x2F;data.bls.gov&#x2F;cgi-bin&#x2F;cpicalc.pl?cost1=279&amp;year1=195701&amp;year2=202406" rel="nofollow">https:&#x2F;&#x2F;data.bls.gov&#x2F;cgi-bin&#x2F;cpicalc.pl?cost1=279&amp;year1=1957...</a>

I never really liked Zuck due to Facebook&#x27;s massive black hole of user tracking, but you&#x27;ve gotta give it to him: the Llama models are insanely good, and the commitment to making Horizon OS available to other VR manufacturers is commendable, although I&#x27;d like to see it truly open sourced.

Ok, but please don&#x27;t post repetitive comments or recite speeches on HN. We&#x27;re trying to avoid that kind of thing here, because repetition is tedious and bad for curiosity.<a href="https:&#x2F;&#x2F;hn.algolia.com&#x2F;?dateRange=all&amp;page=0&amp;prefix=false&amp;sort=byDate&amp;type=comment&amp;query=curiosity%20repetition%20by:dang" rel="nofollow">https:&#x2F;&#x2F;hn.algolia.com&#x2F;?dateRange=all&amp;page=0&amp;prefix=false&amp;so...</a><a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;newsguidelines.html">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;newsguidelines.html</a>

On every post like this on hacker news, I post my pro meta speech. My opinion has always been the same:1. Meta pushed engineering wages higher across the industry.2. They promote high performing engineers very quickly. There are engineers making 7 figures there with just a few years experience.3. They have open sourced the most important frameworks: React and PytorchThis company is a guiding light forcing the hand of other large corporations. Mark Zuckerberg is a hero, and has done a fantastic job

The article argues that the threat of foreign espionage is not solved by closing models.&gt; Some people argue that we must close our models to prevent China from gaining access to them, but my view is that this will not work and will only disadvantage the US and its allies. Our adversaries are great at espionage, stealing models that fit on a thumb drive is relatively easy, and most tech companies are far from operating in a way that would make this more difficult. It seems most likely that a world of only closed models results in a small number of big companies plus our geopolitical adversaries having access to leading models, while startups, universities, and small businesses miss out on opportunities.

This argument implies that cheap phones are bad since telemarketers can use them.

You guys really need to get over your bellicose POV of the world. Actually, before it destroys you. Really, is not necessary. Most people in the world just want to leave in peace, and see their children grow happily. For each data center NK would create there will be a thousand of peaceful, kind, and well-intentioned AI projects going on. Or maybe more.

so that North Korea will create small call centers for cheaper, since they can get these models for free?

This isn&#x27;t what Meta releases with their models, though I would like to see more public training data. However, I still don&#x27;t think that would qualify as &quot;open source&quot;. Something isn&#x27;t open source just because its reproducible out of composable parts. If one, very critical and system defining part is a binary (or similar) without publicly available source code, then I don&#x27;t think it can be said to be &quot;open source&quot;. That would be like saying that Windows 11 is open source because Windows Calculator is open source, and its a component of Windows.

They can&#x27;t release training dataset if it was illegally scrapped all over the web without permission :) (taps head)

Open training dataset + open steps sufficient to train exactly the same model.

You can fine tune the models, and you can modify binaries. However, there is no human readable &quot;source&quot; to open in either case. The act of &quot;fine tuning&quot; is essentially brute forcing the system to gradually alter the weights such that loss is reduced against a new training set. This limits what you can actually do with the model vs an actual open source system where you can understand how the system is working and modify specific functionality.Additionally, models can be (and are) fine tuned via APIs, so if that is the threshold required for a system to be &quot;open source&quot;, then that would also make the GPT4 family and other such API only models which allow finetuning open source.

Can’t you do fine tuning on those binaries?  That’s a modification.

I think you’re right from the perspective of an individual developer. You and I are not about to fork Chromium any time soon. If you presume that forking is impractical then sure, the right to fork isn’t worth much.But just because a single developer couldn’t do it doesn’t mean it couldn’t be done. It means nobody has organized a large enough effort yet.For something like a browser, which is critical for security, you need both the organization and the trust. Despite frequent criticism, Mozilla (for example) is still considered pretty trustworthy in a way that an unknown developer can’t be.

I also think that something like Chromium is a better analogy for corporate open source models than a grassroots project like Linux is. Chromium is technically open source, but Google has absolute control over the direction of it&#x27;s development and realistically it&#x27;s far too complex to maintain a fork without Googles resources, just like Meta has complete control over what goes into their open models, and even if they did release all the training data and code (which they don&#x27;t) us mere plebs could never afford to train a fork from scratch anyway.

&gt; no one has bothered to release their training dataIf the FOSS community sets this as the benchmark for open source in respect of AI, they&#x27;re going to lose control of the term. In most jurisdictions it would be illegal for the likes of Meta to release training data.

AI2 has released training data in their OLMo model:
<a href="https:&#x2F;&#x2F;blog.allenai.org&#x2F;hello-olmo-a-truly-open-llm-43f7e7359222" rel="nofollow">https:&#x2F;&#x2F;blog.allenai.org&#x2F;hello-olmo-a-truly-open-llm-43f7e73...</a>

None of Meta&#x27;s models are &quot;open source&quot; in the FOSS sense, even the latest Llama 3.1. The license is restrictive. And no one has bothered to release their training data either.This post is an ad and trying to paint these things as something they aren&#x27;t.

LLAMA is an open-weights model. I like this term, let&#x27;s use that instead of open source.

&quot;Open weights&quot; means you can use the weights for free (as in beer). &quot;Open source&quot; means you get the training dataset and the methodology. ~Nobody does open source LLMs.

If you can’t share the dataset, under what twisted reality are you fine to share the derivative models based on those unsharable datasets?In a better world, there would be no “I ran some algos on it and now it’s mine” defense.

latest llama 3.1 is in a different repo, <a href="https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama-models&#x2F;blob&#x2F;main&#x2F;models&#x2F;llama3_1&#x2F;api&#x2F;model.py">https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama-models&#x2F;blob&#x2F;main&#x2F;models&#x2F;...</a> , but yes, the code is shared.
It astonishing that in software 2.0 era, powerful applications like llama has only hundreds of lines of code, and most work hidden in training data.
Source code alone is no longer that informative as Software 1.0

Open training data would be great too.If you have open data and open source code you can reproduce the weights

Really? I have to check out the training code again. Last time I looked the training and inference code were just example toys that were barely usable.Has that changed?

For models of this size, the code used to train them is going to be very custom to the architecture&#x2F;cluster they are built on. It would be almost useless to anybody outside of Meta. The dataset would be more a lot more interesting, as it would at the very least show everybody how they got it to behave in certain ways.

&quot;Open weights&quot; is a more appropriate term but I&#x27;ll point out that these weights are also largely inscrutable to the people with the code that trained it. And for licensing reasons, the datasets may not be possible to share.There is still a lot of modifying you can do with a set of weights, and they make great foundations for new stuff, but yeah we may never see a competitive model that&#x27;s 100% buildable at home.Edit: mkolodny points out that the model code is shared (under llama license at least), which is really all you need to run training <a href="https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama3&#x2F;blob&#x2F;main&#x2F;llama&#x2F;model.py">https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama3&#x2F;blob&#x2F;main&#x2F;llama&#x2F;model.p...</a>

That&#x27;s not the training code, just the inference code. The training code, running on thousands of high-end H100 servers, is surely much more complex. They also don&#x27;t open-source the dataset, or the code they used for data scraping&#x2F;filtering&#x2F;etc.

That is just the inference code. Not training code or evaluation code or whatever pre&#x2F;post processing they do.

This is like saying any python program is open source because the python runtime is open source.Inference code is the runtime; the code that runs the model. Not the model itself.

No, it&#x27;s not. The Llama 3 Community License Agreement is not an open source license. Open source licenses need to meet the criteria of the only widely accepted definition of &quot;open source&quot;, and that&#x27;s the one formulated by the OSI [0]. This license has multiple restrictions on use and distribution which make it not open source. I know Facebook keeps calling this stuff open source, maybe in order to get all the good will that open source branding gets you, but that doesn&#x27;t make it true. It&#x27;s like a company calling their candy vegan while listing one its ingredients as pork-based gelatin. No matter how many times the company advertises that their product is vegan, it&#x27;s not, because it doesn&#x27;t meet the definition of vegan.[0] - <a href="https:&#x2F;&#x2F;opensource.org&#x2F;osd" rel="nofollow">https:&#x2F;&#x2F;opensource.org&#x2F;osd</a>

Llama’s code is open source: <a href="https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama3&#x2F;blob&#x2F;main&#x2F;llama&#x2F;model.py">https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama3&#x2F;blob&#x2F;main&#x2F;llama&#x2F;model.p...</a>

They don&#x27;t &quot;[allow] developers to modify its code however they want&quot;, which is a critical component of &quot;open source&quot;, and one that Meta is clearly trying to leverage in branding around its products. I would like them to start calling these &quot;public weight models&quot;, because what they&#x27;re doing now is muddying the waters so much that &quot;open source&quot; now just means providing an enormous binary and an open source harness to run it in, rather than serving access to the same binary via an API.

Ok call it Open Weights then if the dictionary definitions matter so much to you.The actual point that matters is that these models are available for most people to use for a lot of stuff, and this is way way better than what competitors like OpenAI offer.

OSI made themselves the authority because they hated Richard Stallman and his Free Software movement. It&#x27;s just marketing.

Open Source Initiative (kind of a de-facto authority on what&#x27;s open source and what not) is spending a whole lot of time figuring out what it means for an AI system to be open source. In other words, they&#x27;re basically trying to come up with a new license because the existing ones can&#x27;t easily apply.I believe this is the current draft: <a href="https:&#x2F;&#x2F;opensource.org&#x2F;deepdive&#x2F;drafts&#x2F;the-open-source-ai-definition-draft-v-0-0-8" rel="nofollow">https:&#x2F;&#x2F;opensource.org&#x2F;deepdive&#x2F;drafts&#x2F;the-open-source-ai-de...</a>

I think it would also include the code used to train it

You can find the entire Llama 3.0 pretraining set here: <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;datasets&#x2F;HuggingFaceFW&#x2F;fineweb" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;datasets&#x2F;HuggingFaceFW&#x2F;fineweb</a>15T tokens, 45 terrabytes. Seems fairly open source to me.

No. The text is an asset used by the source to train the model. The source can process arbitrary text. Text is just text, it was written for communication purposes, software (defined by source code) processes that text in a particular way to train a model.

&gt; If so, then how can current ML models be open source?The source of a language model is the text it was trained on. Llama models are not open source (contrary to their claims), they are open weight.

I believe it is able to reconstruct parts of the original source material—if the interrogator already knows the original source material to prompt the model appropriately.

One counterpoint is that
major publications (eg New York Times) would have you believe that AI is a mildly lossy compression algorithm capable of reconstructing the original source material.

I dunno — if an open source project required, say, a proprietary compiler, that would diminish its open source-ness. But I agree it&#x27;s not totally comparable, since the weights are not particularly analogous to machine code. We probably need a new term. Open Weights.

Training code is only useful to people in academia, and the closest thing to &quot;code you can modify&quot; are open weights.People are framing this as if it was an open-source hierarchy, with &quot;actual&quot; open-source requiring all training code to be shared. This is not obvious to me, as I&#x27;m not asking people that share open-source libraries to also share the tools they used to develop them. I&#x27;m also not asking them to share all the design documents&#x2F;architecture discussion behind this software. It&#x27;s sufficient that I can take the end result and reshape it in any way I desire.This is coming from an LLM practitioner that finetunes models for a living; and this constant debate about open-source vs open-weights seems like a huge distraction vs the impact open-sourcing something like Llama has... this is truly a Linux-like moment. (at a much smaller scale of course, for now at least)

I think saying it&#x27;s the new binary is closer to the truth. You can&#x27;t reproduce it, but you can use it. In this new version, you can even nudge it a bit to do something a little different.New stuff, so probably not good to force old words, with known meanings, onto new stuff.

I like the term &quot;open weights&quot;. Open source would be the dataset and code that generates these weights.There is still a lot you can do with weights, like fine tuning, and it is arguably more useful as retraining the entire model would cost millions in compute.

Not sure what you mean by &quot;they are good to be a user only.&quot;
Whatever their strategy is, this is great for the community.

If that were the definition then any software you can install on your computer would be open source. It makes open source lose nearly all meaning.Just say &quot;open weights&quot;, not &quot;open source&quot;.

Open source = reproducible binaries (weights) by you on your computer, IMO.Strategy of FB is that they are good to be a user only and fine ruining competitor’s business with good enough free alternatives while collecting awards as saviors of whatever.

If you think about LLMs as a new kind of programming runtime, the matrices are the source.

I think you&#x27;re misunderstanding what I&#x27;m saying. I don&#x27;t think its technically feasible for current models to be open source, because there is no source code to open. Yes, there is a harness that runs the model, but the vast, vast amount of instructions are contained in the model weights, which are akin to a compiled binary.If we make large strides in interpretability we may have something resembling source code, but we&#x27;re certainly not there yet. I don&#x27;t think the solution to that problem should be to change the definition of open source and pretend the problem has been solved.

It&#x27;s no secret that implementing AI usually involves far more investment into training and teaching than actual code. You can know how a neural net or other ML model works. You can have all the code before you. It&#x27;s still a huge job (and investment) to do anything practical with that. If Meta shares the code their AI runs on with you, you&#x27;re not going to be able to do much with it unless you make the same investment in gathering data and teaching to train that AI. That would probably require data Meta won&#x27;t share. You&#x27;d effectively need your own Facebook.If everyone open sources their AI code, Meta can snatch the bits that help them without much fear of helping their direct competitors.

Unfortunately open source really just means an open API these days. The API is heavily intertwined with closed source.

Coming up with the words and concepts to describe the models is a challenge.Does the training data require permission from the copyright holder to use? Are the weights really open source or more like compiled assembly?

Of course you are right, I&#x27;d put it less carefully: The quoted Linux line is deceptive marketing.- If we start with the closed training set, that is closed and stolen, so call it Stolen Source.- What is distributed is a bunch of float arrays. The Llama architecture is published, but not the training or inference code. Without code there is no open source. You can as well call a compiler book open source, because it tells you how to build a compiler.Pure marketing, but predictably many people follow their corporate overlords and eagerly adopt the co-opted terms.Reminder again that FB is not releasing this out of altruism, but because they have an existing profitable business model that does not depend on generated chats. They probably do use it internally for tracking and building profiles, but that is the same as using Linux internally, so they release the weights to destroy the competition.Isn&#x27;t price dumping an anti trust issue?

The term “source code” can mean many things. In a legal context it’s often just defined as the preferred format for modification. It can be argued that for artificial neural networks that’s the weights (along with code and preferably training data).

I agree; there&#x27;s a lot of muddiness in the term &quot;open source AI&quot;. Earlier this year there was a talk[1] at FOSDEM, titled &quot;Moving a step closer to defining Open Source AI&quot;. It is from someone at the Open Source Initiative. The video and slides are available in the link below[1]. From the abstract:&quot;Finding an agreement on what constitutes Open Source AI is the most important challenge facing the free software (also known as open source) movement. European regulation already started referring to &quot;free and open source AI&quot;, large economic actors like Meta are calling their systems &quot;open source&quot; despite the fact that their license contain restrictions on fields-of-use (among other things) and the landscape is evolving so quickly that if we don&#x27;t keep up, we&#x27;ll be irrelevant.&quot;[1] <a href="https:&#x2F;&#x2F;fosdem.org&#x2F;2024&#x2F;schedule&#x2F;event&#x2F;fosdem-2024-2805-moving-a-step-closer-to-" rel="nofollow">https:&#x2F;&#x2F;fosdem.org&#x2F;2024&#x2F;schedule&#x2F;event&#x2F;fosdem-2024-2805-movi...</a>
defining-open-source-ai&#x2F;

You release all the technology and the training data. Everything that was used to create the model, including instructions.I&#x27;m not sure if facebook has done that

No, open source means that sources are open, typically for inspection, modification etc. Also here it can be considered the case. Likely in order to claim &quot;true open source&quot;, they would have to share dataset? But even this might not be enough for truely open source model? This dataset is nothing but another artifact. So how did they arrive at this dataset, now they have to share pipelines and infra..... the thing is, we have not dealt with llm much, it&#x27;s hard to say what can be considered open source llm just yet, so we use that as metaphore for now

&quot;Eventually though, open source Linux gained popularity – initially because it allowed developers to modify its code however they wanted ...&quot;I find the language around &quot;open source AI&quot; to be confusing. With &quot;open source&quot; there&#x27;s usually &quot;source&quot; to open, right? As in, there is human legible code that can be read and modified by the user? If so, then how can current ML models be open source? They&#x27;re very large matrices that are, for the most part, inscrutable to the user. They seem akin to binaries, which, yes, can be modified by the user, but are extremely obscured to the user, and require enormous effort to understand and effectively modify.&quot;Open source&quot; code is not just code that isn&#x27;t executed remotely over an API, and it seems like maybe its being conflated with that here?

&quot;Very few companies can afford the hardware to run these models privately.&quot;I can run Llama 3 70B on my (64GB RAM M2) laptop. I haven&#x27;t tried 3.1 yet but I expect to be able to run that 70B model too.As for the 405B model, the Llama 3.1 announcement says:&gt; To support large-scale production inference for a model at the scale of the 405B, we quantized our models from 16-bit (BF16) to 8-bit (FP8) numerics, effectively lowering the compute requirements needed and allowing the model to run within a single server node.

This is obviously good news, but __personally__ I feel the open-source models are just trying to catch up with whoever the market leader is, based on some benchmarks.The actual problem is running these models. Very few companies can afford the hardware to run these models privately. If you run them in the cloud, then I don&#x27;t see any potential financial gain for any company to fine-tune these huge models just to catch up with OpenAI or Anthropic, when you can probably get a much better deal by fine-tuning the closed-source models.Also this point:&gt; We need to protect our data. Many organizations handle sensitive data that they need to secure and can’t send to closed models over cloud APIs.First, it&#x27;s ironic that Meta is talking about privacy. Second, most companies will run these models in the cloud anyway. You can run OpenAI via Azure Enterprise and Anthropic on AWS Bedrock.

CrowdStrike just added &quot;Centralized Company Controlled Software Ecosystem&quot; to every risk data sheet on the planet. Everything futureproof is self-hosted and open source.

Indeed, fine-tuning is still possible, but you can only go so far with fine-tuning before you need to completely retrain the model.This is why Silo AI, for example, had to start from scratch to get better support for small European languages.

From what i understand the training data and careful curation of it is the hard part. Everyone wants training data sets to train their own models instead of producing their own.

&gt; you can&#x27;t meaningfully modify them given there is almost no information available about the training data, how they were trained, or how the training data was processed.I was under the impression that you could still fine-tune the models or apply your own RLHF on top of them. My understanding is that the training data would mostly be useful for training the model yourself from scratch (possibly after modifying the training data), which would be extremely expensive and out of reach for most people

Indeed, this is something they changed in the 3.1 version of the license.Regardless, the license [1] still has many restrictions, such as the acceptable use policy [2].[1] <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;meta-llama&#x2F;Meta-Llama-3.1-8B&#x2F;blob&#x2F;main&#x2F;LICENSE" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;meta-llama&#x2F;Meta-Llama-3.1-8B&#x2F;blob&#x2F;mai...</a>[2] <a href="https:&#x2F;&#x2F;llama.meta.com&#x2F;llama3_1&#x2F;use-policy" rel="nofollow">https:&#x2F;&#x2F;llama.meta.com&#x2F;llama3_1&#x2F;use-policy</a>

You are definitely allowed to train other models with these models, you just have to give credit in the name, per the license:&gt; If you use the Llama Materials or any outputs or results of the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include “Llama” at the beginning of any such AI model name.

Note that Meta&#x27;s models are not open source in any interpretation of the term.* You can&#x27;t use them for any purpose. For example, the license prohibits using these models to train other models.
* You can&#x27;t meaningfully modify them given there is almost no information available about the training data, how they were trained, or how the training data was processed.As such, the model itself is not available under an open source license and the AI does not comply with the &quot;open source AI&quot; definition by OSI.It&#x27;s an utter disgrace for Meta to write such a blogpost patting themselves on the back while lying about how open these models are.

&gt;You&#x27;re acting as if commoditizing one&#x27;s complements is either new or reprehensible [1].I&#x27;m acting as if calling on other companies to open source their core product, just because it&#x27;s a complement for you, and acting as if it&#x27;s for the benefit of mankind is disingenuous, which it is.

&gt; if Facebook were selling their AI advances instead of selling the output it produces, they wouldn&#x27;t be advocating for everyone else to open source their stackYou&#x27;re acting as if commoditizing one&#x27;s complements is either new or reprehensible [1].[1] <a href="https:&#x2F;&#x2F;gwern.net&#x2F;complement" rel="nofollow">https:&#x2F;&#x2F;gwern.net&#x2F;complement</a>

&gt;The source-code to Ad tracking platform is useless to users.It&#x27;s absolutely not useless for developers looking to build a competing project.&gt;The supposedly &quot;good&quot; and &quot;open&quot; people like Google or OpenAI, haven&#x27;t given their model weights.Because they&#x27;re monetizing it... the only reason Facebook is giving it away is because it&#x27;s a complement to their core product of selling ads. If they were monetizing it, it would be closed source. Just like their Ads platform...

The source-code to Ad tracking platform is useless to users.At the end, it&#x27;s actually Facebook doing the right thing (though they are known for being evil).It&#x27;s a bit of an irony.The supposedly &quot;good&quot; and &quot;open&quot; people like Google or OpenAI, haven&#x27;t given their model weights.A bit like Microsoft became the company that actually supports the whole open-source ecosystem with GitHub.

&gt;In the early days of high-performance computing, the major tech companies of the day each invested heavily in developing their own closed source versions of Unix.Because they sold the resultant code and systems built on it for money... this is the gold miner saying that all shovels and jeans should be free.Am I happy Facebook open sources some of their code? Sure, I think it&#x27;s good for everyone. Do I think they&#x27;re talking out of both sides of their mouth? Absolutely.Let me know when Facebook opens up the entirety of their Ad and Tracking platforms and we can start talking about how it&#x27;s silly for companies to keep software closed.I can say with 100% confidence if Facebook were selling their AI advances instead of selling the output it produces, they wouldn&#x27;t be advocating for everyone else to open source their stacks.

They cannot legally enforce this, because they don’t have the rights to the content they trained it on. Whoever’s willing to fund that court battle would likely win.There’s a legal precedent that says hard work alone isn’t enough to guarantee copyright, i.e. it doesn’t matter that it took millions of dollars to train.

i think these clauses are unenforceable. it&#x27;s telling that OAI hasn&#x27;t tried a similar suit despite multiple extremely well-known cases of competitors training on OAI outputs

Another case of &quot;open-washing&quot;. Llama is not available open source, under the common definition of open source, as the license doesn&#x27;t allow for commercial re-use by default [0].They provide their model, with weights and code, as &quot;source available&quot; and it looks like they allow for commercial use until a 700M monthly subscriber cap is surpassed. They also don&#x27;t allow you to train other AI models with their model:&quot;&quot;&quot;
...
v. You will not use the Llama Materials or any output or results of the Llama Materials to improve any other large language model (excluding Meta Llama 3 or derivative works thereof). 
...
&quot;&quot;&quot;[0] <a href="https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama3&#x2F;blob&#x2F;main&#x2F;LICENSE">https:&#x2F;&#x2F;github.com&#x2F;meta-llama&#x2F;llama3&#x2F;blob&#x2F;main&#x2F;LICENSE</a>

Without Meta, you would still have Mistral, Silo AI, and the many other companies and labs producing much more open models with similar performance.

Everyone complaining about not having data access: Remember that without meta you would have openai and anthropic and that&#x27;s it. I&#x27;m really thankful they&#x27;re releasing this, and the reason they can&#x27;t release the data is obvious.

The irony of this letter being written by Mark Zuckerburg at Meta, while OpenAI continues to be anything but open, is richer than anyone could have imagined.

You can also wait a bit for someone to upload quantized variants, finetunes, etc, and download those. FWIW I&#x27;m not making a claim about the legality of that, just saying it&#x27;s an easy way around needing to sign the agreement.

I think you can use an HF account as well <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;meta-llama" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;meta-llama</a>

It doesn&#x27;t require an account.  You do have to fill in your name and email (and birthdate, although it seems to accept whatever you feed it.)

&quot;open source AI&quot; ... &quot;open&quot; ... &quot;open&quot; ....And you can&#x27;t even try it without an FB&#x2F;IG account.Zuck will never change.

Open source requires, at the very least, that you can use it for any purpose. This is not the case with Llama.The Llama license has a lot of restrictions, based on user base size, type of use, etc.For example you&#x27;re not allowed to use Llama to train or improve other models.But it goes much further than that. The government of India can&#x27;t use Llama because they&#x27;re too large. Sex workers are not allowed to use Llama due to the acceptable use policy of the license. Then there is also the vague language probibiting discrimination, racism etc.. good luck getting something like that approved by your legal team.

GPL defines the “source code” of a work as the preferred form of the work for making modifications to it. If Meta released a petabyte of raw training data, would that really be easier to extend and adapt (as opposed to fine-tuning the weights)?

No, I would argue that from the three main ingredients - training data, model source code and weights - weights are the furthest away from something akin to source code.They&#x27;re more like obfuscated binaries. When it comes to fine-tuning only however things shift a little bit, yes.

No open weights are the output of a proprietary and secretive process of training. It’s like sharing a pre compiled application instead of what you need to reproduce the compiled application.AI2’s OLMo is an example of what open source actually looks like for LLMs:<a href="https:&#x2F;&#x2F;blog.allenai.org&#x2F;hello-olmo-a-truly-open-llm-43f7e7359222" rel="nofollow">https:&#x2F;&#x2F;blog.allenai.org&#x2F;hello-olmo-a-truly-open-llm-43f7e73...</a>

Honest question.  As far as LLMs are concerned, isn&#x27;t open weights same as open source?

&gt; training code and training data would be worthless to anyone who doesn&#x27;t have compute on the level that Meta doesI don&#x27;t fully agree.Isn&#x27;t that like saying *nix being open source is worthless unless you&#x27;re planning to ship your own Linux distro?Knowing how the sausage is made is important if you&#x27;re an animal rights activist.

<a href="https:&#x2F;&#x2F;llama.meta.com&#x2F;llama3_1&#x2F;use-policy&#x2F;" rel="nofollow">https:&#x2F;&#x2F;llama.meta.com&#x2F;llama3_1&#x2F;use-policy&#x2F;</a>The acceptable use policy is seems fine. Don&#x27;t use it to break the law, solicit sex, kill people, or lie.

Meta could change the license of future releases of Llama and kill your business built on it.If the training data was openly available, even if you can&#x27;t afford to res
train a new version, a competitor like Amazon could do it for you

Source versus weights seems like a really pedantic distinction to make. As you say, the training code and training data would be worthless to anyone who doesn&#x27;t have compute on the level that Meta does. Arguably, the weights are source code interpreted by an inference engine, and realistically it&#x27;s the weights that someone is going to want to modify through fine-tuning, not the original training code and data.The far more important distinction is &quot;open&quot; versus &quot;not open&quot;, and I disagree that we should cede that distinction while trying to fight for &quot;source&quot;. The Llama license is restrictive in a number of ways (it incorporates an entire acceptable use policy) that make it most definitely not &quot;open&quot; in the customary sense.

It&#x27;s a real shame that we&#x27;re still calling Llama &quot;open source&quot; when at best it&#x27;s &quot;open weights.&quot;Not that anyone would go buy 100,000 H100s to train their own Llama, but words matter. Definitions matter.

All correct, Meta does obviously benefit.It&#x27;s helpful to also look at what do the developers and companies (everyone outside of top 5&#x2F;10 big tech companies) get out of this. They get open access to weights of SOTA LLM models that take billions of dollars to train and 10s of billions a year to run the AI labs that make these. They get the freedom to fine tune them, to distill them, and to host them on their own hardware in whatever way works best for their products and services.

Meta haven&#x27;t made an open source model. They have released a binary with a proprietary but relatively liberal license. Binaries are not source and their license isn&#x27;t free.

The reason for Meta making their model open source is rather simple: They receive an unimaginable amount of free labor, and their license only excludes their major competitors to ensure mass adoption without benefiting their competition (Microsoft, Google, Alibaba, etc). Public interest, philanthropy, etc are just nice little marketing bonuses as far as they&#x27;re concerned (otherwise they wouldn&#x27;t be including this licensing restriction).

Sure, but don&#x27;t you think the &quot;not getting locked out&quot; is just the pre-requisite for their eventual goal of locking everyone else out?

&gt; I actually think this is one of the rare times where the small guys interests are aligned with MetaSmall guys are the ones being screwed over by AI companies and having their text&#x2F;art&#x2F;code stolen without any attribution or adherence to license. I don’t think Meta is on their side at all

I actually think this is one of the rare times where the small guys interests are aligned with Meta. Meta is scared of a world where they are locked out of LLM platforms, one where OpenAI gets to dictate rules around their use of the platform much like Apple and Google dictates rules around advertiser data and monetization on their mobile platforms. Small developers should be scared of a world where the only competitive LLMs are owned by those players too.Through this lense, Meta’s actions make more sense to me. Why invest billions in VR&#x2F;AR? The answer is simple, don’t get locked out of the next platform, maybe you can own the next one. Why invest in LLMs? Again, don’t get locked out. Google and OpenAi&#x2F;Microsoft are far larger and ahead of Meta right now and Meta genuinely believes the best way to make sure they have an LLM they control is to make everyone else have an LLM they can control. That way community efforts are unified around their standard.

i was thinking along the same. A lot of content generated by LLMs is going to end up on Facebook or Instagram. The easier it is to create AI generated content the more content ends up on those applications.

Especially because genAI is a copyright laundering system. You can train it on copyrighted material and none of the content generated with it are copyright-able, which is perfect for social apps

Meta is, fundamentally, a user-generated-content distribution company.Meta wants to make sure they commoditize their complements: they don’t want a world where OpenAI captures all the value of content generation, they want the cost of producing the best content to be as close to free as possible.

The model it&#x27;s self isn&#x27;t actually that valuable to facebook. The thing that&#x27;s important is the dataset, the infrastructure and the people to make the models.There is still, just about, a strong ethos( especially in the research teams) to chuck loads of stuff over the wall into opensource. (pytorch, detectron, SAM, aria etc)but its seen internally as a two part strategy:1) strong recruitment tool (come work with us, we&#x27;ve done cool things, and you&#x27;ll be able to write papers)2) seeding the research community with a common toolset.

Huge companies like facebook will often argue for solutions that on the surface, seem to be in the public interest.But I have strong doubts they (or any other company) actually believe what they are saying.Here is the reality:- Facebook is spending untold billions on GPU hardware.- Facebook is arguing in favor of open sourcing the models, that they spent billions of dollars to generate, for free...?It follows that companies with much smaller resources (money) will not be able to match what Facebook is doing. Seems like an attempt to kill off the competition (specifically, smaller organizations) before they can take root.

It&#x27;s not cynical, it&#x27;s just an awareness that public companies have a fiduciary duty to their shareholders.

Cynically I think this position is largely due to how they can undercut OpenAI&#x27;s moat.

&gt; the AI folks at Meta deserve our praise and gratitudeWe interviewed Thomas who led Llama 2 and 3 post training here in case you want to hear from someone closer to the ground on the models <a href="https:&#x2F;&#x2F;www.latent.space&#x2F;p&#x2F;llama-3" rel="nofollow">https:&#x2F;&#x2F;www.latent.space&#x2F;p&#x2F;llama-3</a>

Sure, if you have a H100 cluster. If you quant it to int4 you might get away with using only 4 H100 GPUs!

If you have the ram for it.Ollama will offload as many layers as it can to the gpu then the rest will run on the cpu&#x2F;ram.

If you want your first token around tomorrow lunch sure

&gt; We’re releasing Llama 3.1 405BIs it possible to run this with ollama?

All said and done, it is a very expensive and balsy way to undercut competitors. They’ve spent &gt; $5B on hardware alone, much of which will depreciate in value quickly.Pretty sure the only reason Meta’s managed to do this is because of Zuck’s iron grip on the board (majority voting rights). This is great for Open Source and regular people though!

&gt;selling user generated content and private info to the highest bidderWas always their modus operandi, surely. How else would they have survived.Thanks for returning everyone else;s content and never mind all the content stealing your platform did.

I&#x27;m perfectly happy with them draining the life essence out of the people crazy enough to still use Facebook, if they&#x27;re funneling the profits into advancing human progress with AI. It&#x27;s an Alfred Nobel kind of thing to do.

Praising is good. Gratitude is a bit much. They got this big by selling user generated content and private info to the highest bidder. Often through questionable means.Also, the underdog always touts Open Source and standards, so it’s good to remain skeptical when&#x2F;if tables turn.

&quot;Come to think of it, the negative impact on such competitors was likely a key motivation for releasing the new models.&quot;&quot;Commoditize Your Complement&quot; is often cited here: <a href="https:&#x2F;&#x2F;gwern.net&#x2F;complement" rel="nofollow">https:&#x2F;&#x2F;gwern.net&#x2F;complement</a>

If the model is already pretrained, there&#x27;s no need to pretrain it, so the cost of pretraining is zero.

&gt; they have brought the cost of pretraining a frontier 400B+ parameter model to ZEROIt is still far from zero.

Its certainly not altruism. Given that Facebook&#x2F;Meta owns the largest user data collection systems, any advancement in AI ultimately strengthens their business model (which is still mostly collecting private user data, amassing large user datasets, and selling targeting ads).There is a demo video that shows a user wearing a Quest VR headset and asks the AI &quot;what do you see&quot; and it interprets everything around it. Then, &quot;what goes well with these shorts&quot;... You can see where this is going. Wearing headsets with AIs monitoring everything the users see and collecting even more data is becoming normalized. Imagine the private data harvesting capabilities of the internet but anywhere in the physical world. People need not even choose to wear a Meta headset, simply passing a user with a Meta headset in public will be enough to have private data collected. This will be the inevitable result of vision models improvements integrated into mobile VR&#x2F;AR headsets.

You can always listen to the investor calls for the capitalist point of view. In short, attracting talent, building the ecosystem, and making it really easy for users to make stuff they want to share on Meta&#x27;s social networks

I really think the value of this for Meta is content generation. More open models (especially state of the art) means more content is being generated, and more content is being shared on Meta platforms, so there is more advertising revenue for Meta.

All the content generated by llms (good or bad) is going to end up back in Facebook&#x2F;Instagram and other social media sites. This enables Meta to show growth and therefore demand a higher stock price. So it makes sense to get content generation tools out there as widely as possible.

He&#x27;s not even pretending it&#x27;s altruism. Literally about 1&#x2F;3 of the entire post is the section titled &quot;Why Open Source AI Is Good for Meta&quot;. I find it really weird that there are whole debates in threads here about whether it&#x27;s altruistic when Zuckerberg isn&#x27;t making that claim in the first place.

He addresses this pretty clearly in the post. They don&#x27;t want to be beholden to other companies to build the products they want to build. Their experience being under Apple&#x27;s thumb on mobile strongly shaped this point of view.

Zuckerberg didn&#x27;t really say anything about altruism. The point he was making is an explicit &quot;I believe open models are best for our business&quot;He was clear in that one of their motivations is avoiding vendor lockin. He doesn&#x27;t want Meta to be under the control of their competitors or other AI providers.He also recognizes the value brought to his company by open sourcing products. Just look at React, PyTorch, and GraphQL. All industry standards, and all brought tremendous value to Facebook.

Makes me wonder why he&#x27;s really doing this. Zuckerberg being Zuckerberg, it can&#x27;t be out of any genuine sense of altruism. Probably just wants to crush all competitors before he monetizes the next generation of Meta AI.

There&#x27;s nothing open source about it.It&#x27;s a proprietary dump of data you can&#x27;t replicate or verify.What were the sources? What datasets it was trained on? What are the training parameters? And so on and so on

&gt; If that ever stops being the case, you should fork rather than trust their judgment.Fork what? The secret sauce is in the training data and infrastructure. I don&#x27;t think either of those is currently open.

Is forking really possible with an LLM or one the size of future Lama versions, have they even released the weights and everything? Maybe I am just negative about it because I feel Meta is the worst company ever invented and feel this will hurt society in the long run just like Facebook.

In tech you can trust the underdogs. Once they turn into dominant players they turn evil. 99% of the cases.

Look, absolutely zero people in the world should trust any tech company when they say they care about or will keep commitments to the open-source ecosystem in any capacity. Nevertheless, it is occasionally strategic for them to do so, and there can be ancillary benefits for said ecosystem in those moments where this is the best play for them to harm their competitorsFor now, Meta seems to release Llama models in ways that don&#x27;t significantly lock people into their infrastructure. If that ever stops being the case, you should fork rather than trust their judgment. I say this knowing full well that most of the internet is on AWS or GCP, most brick and mortar businesses use Windows, and carrying a proprietary smartphone is essentially required to participate in many aspects of the modern economy. All of this is a mistake. You can&#x27;t resist all lock-in. The players involved effectively run the world. You should still try where you can, and we should still be happy when tech companies either slip up or make the momentary strategic decision to make this easier

They open sourced Gear VR. The are a stark contrast to other players in terms of how they have built everything on open standards (OpenXR, WebXR, etc), and they have just opened their platform by allowing third parties to build and customise it to make their own commercial offerings. Not open source, but a quite a contrast to every other player in that industry so far.

&gt;&gt; Bravo! While I don&#x27;t agree with Zuck&#x27;s views and actions on many fronts, on this occasion I think he and the AI folks at Meta deserve our praise and gratitude.Nope. Not one bit. Supporting F&#x2F;OSS when it suits you in one area and then being totally dismissive of it in every other area should not be lauded. How about open sourcing some of FB&#x27;s VR efforts?

&gt; We’re releasing Llama 3.1 405B, the first frontier-level open source AI model, as well as new and improved Llama 3.1 70B and 8B models.Bravo! While I don&#x27;t agree with Zuck&#x27;s views and actions on many fronts, on this occasion I think he and the AI folks at Meta deserve our praise and gratitude. With this release, they have brought the cost of pretraining a frontier 400B+ parameter model to ZERO for pretty much everyone -- well, everyone except Meta&#x27;s key competitors.[a] THANK YOU ZUCK.Meanwhile, the business-minded people at Meta surely won&#x27;t mind if the release of these frontier models to the public happens to completely mess up the AI plans of competitors like OpenAI&#x2F;Microsoft, Google, Anthropic, etc. Come to think of it, the negative impact on such competitors was likely a key motivation for releasing the new models.---[a] The license is not open to the handful of companies worldwide which have more than 700M users.

Open source AI is the path forward