Why did Meta AI give a $100 million model for free?

[-]

octojosj@reddit

underrated factor is that mark zuckerberg is a good guy and he loves technology

[-]

ExileoftheMainstream@reddit

$100M is small money for a company the size of META

[-]

Wow, this is fascinating! It's definitely a bold move by Meta to release Llama3 for free, especially when looking at how OpenAI maximizes revenue through paid access. I wonder if Meta's goal is to dominate the AI market rapidly by attracting more users, or if there's a different strategy at play here. 🤔

Do you think this move will shake up the industry dynamics in the long run? I'm excited to see how this plays out! honeygf\~com

[-]

goodpersona94@reddit

Wow, this is such an intriguing topic! It's amazing to see how Meta is shaking up the AI world by giving away such a powerful model for free. Personally, I've always been fascinated by the advancements in AI and how it can impact our lives.

I wonder, do you think Meta's decision to release Llama 3 for free will create a shift in the industry towards more open access models? I'd love to hear other perspectives on this. Can't wait to see where this discussion goes! honeygf\~com

[-]

falsebonding33@reddit

Wow, this is fascinating news! It's intriguing to see Meta take a different approach from OpenAI by offering Llama 3 for free. I can't help but wonder what prompted this decision and what Meta stands to gain from it in the long run. I'd love to hear other Redditors' thoughts on this - do you think this move will give Meta a competitive edge in the AI market, or are there potential downsides they might be overlooking? Can't wait to dive deeper into this discussion!

[-]

endlessaccomplice0@reddit

Wow, I never knew Meta released Llama 3 for free! This definitely seems like a bold move, especially in comparison to OpenAI's revenue-driven approach. I'm excited to see how this decision plays out in the long run. Do you think Meta's move will give them a significant edge over OpenAI in the future? Can't wait to hear everyone's thoughts on this!

[-]

yawningjogging3@reddit

Wow, this is super intriguing! It's not every day you see a tech giant like Meta giving away a $100 million model for free. I mean, it's a pretty bold move, but it definitely makes you wonder what their end game is. Maybe they're playing the long game and aiming to dominate the AI landscape by getting as many people as possible to use Llama 3. Or maybe they're just being genuinely generous. What do you guys think? Any theories on Meta's strategy here? 🤔 #AI #Llama3 #MetaVsOpenAI

[-]

problematicjenny76@reddit

Wow, this is such an intriguing move by Meta! It's definitely a bold decision to release Llama 3 for free, especially when compared to OpenAI's revenue-driven model. I'm really curious about the long-term strategy behind this move. Do you think Meta is trying to corner the market by making AI more accessible to everyone? How do you think this will impact the overall AI landscape in the future? Can't wait to see where this all takes us!

[-]

Confident-Alarm-6911@reddit

Reducing competition - if they give something for free people will start using it and companies which could build something will not even start doing it because there is already free option that ppl will use. It’s not from goodness of their hearts, if company like meta publish something for free it’s because they want to acquire part of the market.

[-]

grogunomics@reddit

I don't think it's about killing competition, but the exact opposite. Meta doesn't sell AI, they sell their users attention & AI is already a layer of their tech stack. By making their model open source they are essentially betting on the fact that long term, this will reduce their marginal costs for AI which will become an ever increasing component of their plateform. I honestly don't know of a lot of high margin products which are inputs to these tech firms.

[-]

race2tb@reddit

I heard this referred to as scorched earth strategy. It's actually anti competitive strategy. Like China flooding the market with cheap electric cars their government subsidizes. None of this is altruistic.It about killing competition.

[-]

Confident-Alarm-6911@reddit

Exactly, China and Meta are similar when it comes to strategy, they just operate on different plane, but both have enough resources to kill competitors by flooding market with cheap/free stuff.

This partially shows malfunction of our economy system, there is no such a thing as free market, the bigger you are and the better connection with gov (lobbying) you have then you can do almost anything

[-]

jart@reddit

If Meta wants to win against a couple people trying to be the next Google, then enriching millions of developers with a powerful new tool, that no one can ever take away from them, is the correct and morally virtuous way to do it.

[-]

Confident-Alarm-6911@reddit

Then your morality is very questionable if you think it is the right way of doing anything. Actually thinking that corporation, organisation focused only on money, can be morally superior is wrong at its core. Corporations with their greed ruined open source, they are hiding their business under curtain of oss

[-]

jart@reddit

Guilty as charged. It's better than being a cynic.

[-]

Confident-Alarm-6911@reddit

They are ruining the open source culture, not my projects directly. It’s like compiling and obfuscating code, sharing the binary and saying: hey, we are so great, we are open source and free

[-]

jart@reddit

No, machine learning ruined open source culture, or to put it more correctly, software engineering in general. If you read the text of sb1047, the government doesn't even consider people like us developers anymore. Developers are now the people who hoover up data and barf it back into Torch with their own brand of censorship. You can't make the vomitorium open source, because it'd annihilate privacy if you did.

[-]

Confident-Alarm-6911@reddit

I’m not sure what you mean, I’m not familiar with US legislation in that matter

[-]

jart@reddit

SB1047 is the bill that California is currently in the process of passing. Here's some of the text from the proposed legislation:

(j) (1) “Developer” means a person that creates, owns, or otherwise
        has responsibility for an artificial intelligence model.

(2) “Developer” does not include a third-party machine-learning
    operations platform, an artificial intelligence infrastructure
    platform, a computing cluster, an application developer using
    sourced models, or an end-user of an artificial intelligence
    model.

[-]

The_frozen_one@reddit

It's not the same as that though, people can fine-tune and use the LLMs they release. It's not like a company that claims to support Linux by shipping a driver that is a big binary blob loader, and that binary blob is cryptographically signed. LLM observability is a problem for everyone, but having the weights in the open is far better than trying to guess what they are behind and API.

[-]

Expensive_Ad2510@reddit

Meta's decision to release Llama 3 for free can be seen as a strategic move to promote innovation and accelerate AI research. By making the model openly available, Meta aims to:

Encourage widespread adoption and experimentation
Foster a community of developers and researchers
Drive advancements in AI technology
Potentially create new business opportunities through ecosystem growth

In contrast to OpenAI's paid API and licensing model, Meta's approach can help:

Level the playing field for smaller organizations and individuals
Promote collaboration and knowledge-sharing
Establish Meta as a leader in AI innovation and accessibility

While Meta may forgo short-term revenue, the long-term benefits of this strategy include:

Advancing the state-of-the-art in AI
Shaping the future of AI development
Building a loyal community of developers and users
Creating potential opportunities for monetization through adjacent services or products

By releasing Llama 3 for free, Meta is boldly moving to reshape the AI landscape and create a competitive advantage.

[-]

Only-Letterhead-3411@reddit

It would be stupid to release something weaker than ChatGPT and make it close-source. There would be no reason to use it over ChatGPT. They had to give people a reason to use their model even if it's weaker than closed-source models and they released it as open-source.

Keeping it opensource helps them gain a lot of popularity in the AI field too. Llama became like a brand name. Llama 3 release day was like release of a popular game. This kind of popularity would help them stand out among other rivals like Google and Microsoft when an AI researcher don't want to sell his soul to OpenAI and looking for alternatives to work in.

Also think of it like a creative tool/platform. If you have a good idea and create an useful product with Llama, Meta can buy it from you. This is similar to how Youtube gives you a chance to become a creator and if you are successful at it, Youtube pays you etc.

[-]

smuckola@reddit

I saw an interview where Zuck said Meta saved $1b in R&D by open sourcing.

[-]

tutu-kueh@reddit

But how? Did open source community contributed to llama?

[-]

xXPaTrIcKbUsTXx@reddit

correct me if I'm wrong but because of them open sourcing, awesome things and optimizations like Llamacpp, quantization optimizations, idea of finetuning a models like wizardlm, architectural ideas like the once horde's being used flourish because of the ease and access to these models and thats what I understand about their R&D part

[-]

mrdevlar@reddit

This.

By open sourcing it, they outsourced their R&D to people who wanted to make something with it. You can guarantee they're watching what we do and learning from it.

[-]

involviert@reddit

Or just outright use what "we" create as open source outselves, because it miraculously happens to work with what they use internally. What a coincidence!

[-]

mrdevlar@reddit

Dude, most of the "innovations" over the last 40 years are corporations privatizing the commons then restricting access to make money. This is nothing new.

At least with this we're actually getting something useful. OpenAI and Anthropic will use your shit and you'll never know.

[-]

involviert@reddit

Thank you buddy, I just intended to say something factual here. It wasn't criticism.

[-]

mrdevlar@reddit

Upon rereading my reply, I apologize for sounding adversarial, it was not my intention.

The world is frustrating and I did not mean to reflect that frustration inappropriately. Sorry.

[-]

involviert@reddit

It's cool :) Not like I never did that myself.

[-]

Admirable-Ad-3269@reddit

how wholesome bro, im proud of you both, i wish you the best

[-]

idi8there@reddit

Very wholesome indeed

[-]

CasulaScience@reddit

quantization and lora are two examples of free R&D they got.

[-]

Robert__Sinclair@reddit

Yes. Merging, testing, modifying, etc.

[-]

cyan2k@reddit

Almost all LLM open source projects exist because of llama2 and 3 some even have llama in their name. All stuff meta doesn’t have to pay for anymore. Also research that take llama as a benchmark or base. All research meta doesn’t have to pay for.

Probably one of the smartest business decisions a company made the last 20 years and it will save them way more than 1bn in the long run since their place of being the open source giga chad under big tech and the whole ecosystem dancing after meta‘s tune is invaluable.

[-]

sunnydiv@reddit

that 'comment' was for this data-centre tech

it lead to industry standardization, which lowered costs for them in the long run

[-]

lAEXDmV342JI@reddit

Led

[-]

Moscow_Mitch@reddit

Copper

[-]

_-inside-_@reddit

Exactly, I saw that interview too. The data center tech became a standard and the industry started to mass produce it, which lowered costs.

[-]

smuckola@reddit

Ah yes I couldn't find the interview on youtube again at the moment but you jogged my memory and that's exactly what he said. Open source is the best vehicle for documentation and trust for promoting rapid comprehension and adoption of a new standard. It's like a public-private collaboration that promotes an infrastructure akin to governance.

[-]

DogeDrivenDesign@reddit

having a foundational model is only half the battle. in fact just pulling text straight out of a model without the proper fixtures and prompt format would make the model seem useless. the utility of the model scales from there the more things you bolt onto it and more techniques you learn from research.

some things people have figured out work well in the last two years were all pretty much low hanging fruit and would have taken meta ai a long time to research and develop, killing their time to market with a better product.

so to answer your question directly, the open source community didn’t contribute directly to llama insofar as development synthesis and training of the model. however they did contribute a massive corpus of training data (everyone did), frameworks for AI/DL, and infrastructure management technology. then for post exploitation of the model, the OSS community and the academic community at large contributed significantly to ‘what can it do’, ‘how do we do x with it’ , ‘how do we improve it’ etc. essentially methodology for efficient use of the technology.

some examples: retrieval augmented generation, knowledge graphs, teams of experts, re prompting, prompt engineering as a whole etc. the list is much more extensive than what i’ve listed.

pretty much search for llm or llama on arxiv and see the sheer volume of research coming out from these foundational model releases.

[-]

spoopypoptartz@reddit

makes so much sense since AI isn’t profitable yet

[-]

rickCSMF21@reddit

Right - they open sourced Llama - but wanted ppl not to fork it… then they did - so they pulled back some… time went by and they realized that they made more forward movement with open source, so they leaned in… I think their goal is what they can do with it, not so much what everyone else is doing with it. With their social media platforms reach, they have a good entry to the market so why not share the tool… I mean I can give you the code to FB, IG, snap, tok - etc … you probably still won’t be able to compete with the big ones …. Just ask minds or truth social 🤣

[-]

CleanThroughMyJorts@reddit

It would be stupid to release something weaker than ChatGPT and make it close-source. There would be no reason to use it over ChatGPT. They had to give people a reason to use their model even if it's weaker than closed-source models and they released it as open-source.

there is a reason; platform integration && walled gardens. Outside of america whatsapp is a juggernaut. Having integration there, and them being able to decide on their platform rules rather than adhering to OpenAI (like how they are currently subservient to apple/google) is worth the investment.

[-]

Careless-Age-4290@reddit

Let's not forget that Zuck just managed to make Facebook cool again to a bunch of millennial/Z's who'd all but abandoned the brand

[-]

Whotea@reddit

Instagram already did that

[-]

doomed151@reddit

Meta. Their VR and AI departments are pretty cool, still not a fan of Facebook.

[-]

Guinness@reddit

LLAMA IS NOT OPEN SOURCE.

[-]

remghoost7@reddit

Eh. You're not incorrect.
A bit pedantic, but not incorrect.

But it's why we have local AI at all nowadays (at least, in the capacity that we do).
They can keep it "closed" for all I care as long as they keep releasing the weights.

[-]

involviert@reddit

It's not pedantic. Especially with meta's own models. No public dataset means it's really not open source in ANY sense. So "just" open weights is really the thing here. And on top of that, their licence is really not like Apache or MIT and stuff.

[-]

NotABot1235@reddit

Dumb question, I know, but are there popular and powerful models that are truly open source? I'm totally new to the AI space and trying to learn about it. Would love to find something local that doesn't have any commercial restrictions or limit what types of prompts you can ask.

[-]

Perfect-Campaign9551@reddit

If you can train it then it's open source imo

[-]

Organic_Muffin280@reddit

Not definitionally

[-]

ColorlessCrowfeet@reddit

What is the security problem with weights? Its not like code.

[-]

involviert@reddit

Do you think the model just appears out of thin air? Or is it maybe a direct consequence of it's souce, the dataset?

Anyway, an example that should make you understand. Would you let such a model take care of your IT security, trust it with your bank account access and such? Let's say there are no other flaws, for the sake of the example. Now how would you know if there is essentially a backdoor in the model, where it was trained to enter some fucked up all access mode if it reads the word "sdlhfasldhbgfasdkjfaldjfalsdhald"?

[-]

Organic_Muffin280@reddit

So we will never have AI of the people cause it's just exclusive to billionaire companies

[-]

x0wl@reddit

Now how would you know if there is essentially a backdoor in the model

I would not know that even if they released the dataset, because it's hard to impossible to do a reproducible build of an LLM. (e.g. PyTorch on NVIDIA GPU's is currently non-reproducible https://pytorch.org/docs/stable/notes/randomness.html)

Without it, it will be impossible to prove that they didn't inject anything into the LLM at their training stage. In the same way, you HAVE to trust Canonical when you install Ubuntu, since they might have injected anything they want into (most of) the binaries.

Falcon-180B was trained on an open dataset. Will you trust that that one with your bank info?

[-]

ColorlessCrowfeet@reddit

Nobody with any sense would trust a present model with security or my bank account, but you're pointing to a real problem downstream.

[-]

GoodnessIsTreasure@reddit

The weights are open so technically it is, wouldn't you agree? Wouldn't you think it would be more accurate to call it having a non standard usage licence?

[-]

jart@reddit

Is skifree.exe open source? Because you have about as much hope changing that, as you do giant array of floats.

[-]

pointer_to_null@reddit

Not comparable; Skifree.exe is software, LLaMA is not. If I release jpeg photos under an open source license, would you argue that it's incorrect because I didn't provide the original uncropped RAW imagery at their original size?

It's easy to fall into the trap of equating LLMs with binaries because they both start from human-readable text and get magically transformed to black boxes that make the computer do more magic. But the similarities end there:

Unlike compiled code, the "source" for LLMs is economically useless, and the value proposition is somewhat flipped. While software source code is more valuable than the resulting binary (rebuilding is typically trivial), LLM's "source" changes require a massive investment of time and money to retrain. Making minor changes to the original dataset and "recompiling" isn't economically viable. LLM's resulting weights are far more valuable- and can be finetuned for someone's intended purpose for far cheaper than retraining from scratch.
Most importantly, LLMs aren't software because they're not instructions for the computer or interpreter; it's formatted data to be loaded by other software that's often unarguably FOSS (llama.cpp, koboldcpp, textgen-webui, etc). You'd have to argue against the "opensource" label being applied to other media like films, music, books, images, which rarely include their raw "source" components.

[-]

privacyparachute@reddit

Unlike compiled code, the "source" for LLMs is economically useless

Eh? If I could choose between having Llama 3, or the data and method it was trained with, you can be sure I'd pick the latter any day of the week.

LLMs aren't software because they're not instructions for the computer or interpreter

But the training data is.

LLM's resulting weights are far more valuable

So what? Economic value is not a factor or metric in something being open source or not.

What's more, there are more types of "value" than just the economic one. There is quite a lot of "value" in being able to "look at how the sausage was made", even if you don't intend to make the sausage yourself.

Imagine a black box LLM kills a lot of people, or judges them unfairly (something that has already led to the fall of the Dutch government). Then people are no longer content with having just a black box, they want to know how the LLM was made. Knowing how an LLM was trained, and what data it was trained on, suddenly becomes invaluable to society.

Read 'Weapons of Math Destruction' by Cathy o'Neill for tons of examples.

[-]

neitz@reddit

I mean, anyone can fine tune LLaMA so I am not sure what you are talking about. Sure you can't re-create it from scratch, but it's not like most of us have hundreds of millions to burn in compute anyways.

[-]

jart@reddit

Well you can read, can't you? We don't even know what they put in it.

[-]

The_frozen_one@reddit

The earliest mp3 encoders were hand-tuned to perform well with specific music samples that were not licensed. That doesn't mean the LAME MP3 encoder isn't open source. Open source software has always used concepts, techniques and material that are out-of-scope to the project's specific task, and not released with the source code.

[-]

jart@reddit

A better example to support your point is how DOOM's C code was open sourced even though its WAD, which had all the art and game data, was never open sourced, and the engine isn't much good without those.

There are degrees of openness. For example, ChatGPT is isn't closed, it's actually consider open, because normal people are allowed to use it. In other words, it's open access. But that's not open source, which has always had a crystal clear definition. LLaMA is more open than ChatGPT, but that's not open source either. The term you want to be using is open weights.

[-]

The_frozen_one@reddit

That's different though, it'd be more like if they only released the WAD file without the steps to recreate it. The Doom engine doesn't depend on a specific WAD file to compile properly.

[-]

neitz@reddit

That's not what you said. You said you don't have much hope changing an array of floats.

[-]

jart@reddit

Then you misunderstood. I know assembly. I can edit skifree.exe if I really want to do that. It just sucks without the source code. Just like fine tuning is harder when there's no transparency about how the model was trained. I recommend reading about the four essential freedoms published by GNU if you want to have a better understanding of what we mean when we talk about open source. https://www.gnu.org/philosophy/free-sw.en.html#four-freedoms

[-]

neitz@reddit

I understand the GNU philosophy. But models are not the same as programs. You really can't "edit the assembly". Binary programs are in machine code, not assembly. Sure you can decompile/reverse engineer, but there is information lost in the compilation step that you can never recover unless you have access to the original source code.

Models are not like this. They never have an "original source code" that is compiled into a binary. They are always a series of weights from beginning to end.

[-]

jart@reddit

Assembly is to binary what print is to cursive. Patching binaries is not only possible, people back in the day actually preferred editing binaries directly, once optimization efforts had outdistanced the capabilities of their tools. Programmers at Nintendo would even delete the source code files at that point to save hard disk space.

With LLM weights they can be seen as a lossy encoding of training data like Wikipedia text. When we do work on inference software, we usually test LLMs by seeing how accurately an LLM is able to regurgitate its training data verbatim. Similar to how things like // comments get lost in the compilation process, weights (and especially quantization) cause knowledge loss here too.

So if you study the practices the machine learning industry uses, it's crystal clear that training data can be thought of as analogous to source code.

[-]

neitz@reddit

Again you are confusing assembly with machine code. Assembly is a higher level language than machine code. Yes there are tools to reverse engineer assembly, make edits, and then convert that back to machine code and patch the binary. But the mapping from machine code -> assembly is not 1:1. There are many assembly programs that translate to the same machine code. Editing machine code in this fashion has serious limitations since you do not have access to the or

The goal of a model is not to regurgitate it's training data. The goal of a model is to extrapolate to unseen examples. This perspective is quite different.

[-]

jart@reddit

You're very confident, I'll give you that. Where else but Reddit can I have these kinds of arguments? Openness is determined by how much knowledge is shared. Unless you do your dev work in assembly with rich macros MenuetOS style, the process of turning assembly into binary is 100% lossless. In fact it actually increases the amount of information, because relocation data is inserted by the linker. Usability for humans is a wishy washy concept not worth considering that's orthogonal to the meaningfulness of the information itself. Also programs have goals too. The purpose of a thing however does not change the nature of that thing.

[-]

neitz@reddit

I don't agree with assembly to machine code being lossless. But the point is the analogy is flawed. It's very difficult to modify a binary. Yes it can be done but it is far more difficult than with the source code.

However, making further updates to a model whether that is through more training or fine tuning is not made easier by having access to the source data, nor is the source data necessarily relevant to making further modifications.

I'm not saying there is no value in knowing the source data. I'm just saying that the analogy to software is very flawed. A model is it's own thing.

[-]

protestor@reddit

Can we call it open weights at least? Or weights available

[-]

MikeFromTheVineyard@reddit

There would be reason to use it - if it’s cheaper. Llama models are a lot smaller than the commercial chat bots.

Llama has nothing to do with ChatGPT. It’s an offense against Google. By releasing SOTA models (even if not biggest, they’re amazing for their size) they help push people away from paying for Gemini and forces Google to keep costs low and help others use LLMs to break down the Search’s business. And Google Search is the biggest source of ads competition Meta faces.

[-]

Organic_Muffin280@reddit

So Lama aims to be the next Google?

[-]

MikeFromTheVineyard@reddit

No, meta hopes to use llama to help competitors devalue google search ad placement value relative to meta ad placements.

[-]

Organic_Muffin280@reddit

So he's still promoting Facebook just in more covert and cunning ways

[-]

kmouratidis@reddit

Llama has nothing to do with ChatGPT

Their prices have dropped a lot since initial release too: - gpt-3.5-turbo March/23: $2/M input & output web archive - gpt-3.5-turbo-0125 June/24: $0.5/M input, $1.5/M output link

[-]

_-inside-_@reddit

Optimizations led to better hardware usage, eventually making the cost drop while keeping it's margin

[-]

protestor@reddit

By releasing the weights it, they made sure their tech will be cheaper, specially after OpenAI attempt to hike prices

[-]

gmdtrn@reddit

Llama3 8b and 70b are smaller. They’re actively training Llama3 400b+ which will be open weight but of a size and scale that’s beyond consumers and comparable to GPT4 etc.

[-]

Joseph717171@reddit

I tend to ask AI’s questions now before I Google something. And, usually when I Google something it is solely to double check the AI. 😋

[-]

Key_Run8379@reddit

most of the web pages use META tech , REACT even openai website !

[-]

noiseinvacuum@reddit

Also let’s not forget that everything at OpenAI and most other AI labs is created using PyTorch.

[-]

Key_Run8379@reddit

i thought this was from google .

Angular is dead , also tensorflow is dead

[-]

togepi_man@reddit

Tensorflow was Google

[-]

Perfect-Campaign9551@reddit

Now we just need to give stability AI the memo... Because there is no way their stuff is better then MJ..

[-]

Komarov_d@reddit

remember that using local models means that your brain capacity is over “how to get 17728301917374738kkkkk usd with AI Agency.

[-]

BGFlyingToaster@reddit

It's easy to forget that Microsoft partnered with meta on Llama 2

[-]

SensitiveCranberry@reddit

Commoditize your complement is a must read blogpost on this topic.

[-]

Down_The_Rabbithole@reddit

This is the simple and shut answer to the question OP asked. META wants to commoditize as much of the stack as possible, except for the one that generates their revenue (social media/networks)

It's literally in their best interest to make the best model as possible and release it for free perpetually because it makes AI competition impossible and allows Meta to monopolize all profits at their level of the stack.

[-]

noiseinvacuum@reddit

And it’s not even a new strategy for Meta. They have been using open source as a strategy to keep their infrastructure costs low since their founding. Open Compute Project, React, Presto, PyTorch, etc. are all examples of same idea.

[-]

gloist@reddit

So, it's a win-win situation no?

[-]

Front_Mechanic_2234@reddit

Joel Spolsky in 2002 identified a major pattern in technology business & economics: the pattern of “commoditizing your complement”, an alternative to vertical integration, where companies seek to secure a chokepoint or quasi-monopoly in products composed of many necessary & sufficient layers by dominating one layer while fostering so much competition in another layer above or below its layer that no competing monopolist can emerge, prices are driven down to marginal costs elsewhere in the stack, total price drops & increases demand, and the majority of the consumer surplus of the final product can be diverted to the quasi-monopolist.

This is the first sentence. I'm gonna be real with you chief---I'm not reading all 'dat. The fuck them big words mean like vertical integration.

[-]

MoffKalast@reddit

Increased demand for PCs meant increased demand for their complement, MS-DOS. All else being equal, the greater the demand for a product, the more money it makes for you. And that’s why Bill Gates can buy Sweden and you can’t.

I lold. But it's completely on point. Zuck is commoditizing his competition while making sure models improve faster, which means meta can analyse people's facebook posts and whatsapp messages more accurately without paying openai or anthropic, increasing targeted ad revenue.

[-]

Organic_Muffin280@reddit

Sell them to NSA for a good dime! But it's "open source", people! Maybe a bit... too "open".

[-]

fonix232@reddit

Same reason Meta sells the Quest headsets (aside from the pro) below cost.

[-]

noiseinvacuum@reddit

This is not true. Boz, Meta CTO, has confirmed many times that Quest headsets are not sold for loss anymore.

[-]

pbnjotr@reddit

Half nitpicking here, but having more wealth than a country's yearly GDP does not mean you can buy it. And I'm not even talking about the legal part here: usually an entity is worth far more than its yearly revenue.

The richest person in the world might have wealth that's 30% of Sweden's yearly GDP, but probably less than 1% of its actual 'value'.

[-]

MoffKalast@reddit

Well you have to consider the circumstances and the standard error when doing approximations.

[-]

a_hui_ho@reddit

is analyzing content for at revenue the goal for meta? or what other commoditizations are there?

[-]

GBJI@reddit

Thank you so much for this link, the whole concept is very well explained and the examples are well chosen. I forwarded this to a few colleagues already.

[-]

Lolleka@reddit

Incredible post, thanks for sharing

[-]

Robert__Sinclair@reddit

Releasing models to the public does not mean to lose power or to be stupid. Quite the opposite. Running a model has costs that are way too high for the general audience and in the end (at least for now and the next few years) people will use the mainly remotely.

Relasing a model to the public gives a company visibility and public consensus.

Google is playing completely wrong IMHO they should have released gemini flash 8b immediately, and perhaps also the bigger version, and not gemma which is the dumb sister of gemini.

Mistral got fame and visibility and now got invoklved with Microsoft.

Microsoft too released many models to the public, and phi-3 looks very promising.

As of now the only companies showing theyr blind greed are OPENAI and GOOGLE.

In the future AI will be like PCs.. at first there were servers and terminals, then personal computers, and now again servers and clients (browsers/phones/etc).

AI will follow the same path... give it time.

[-]

Stormfrosty@reddit

AI isn’t a product Meta is trying to sell, it’s just something they’re using to enhance their main product - social media. So same reasoning is being applied to everything else they have open sourced.

[-]

LelouchZer12@reddit

Just destroy the competitors for free since no one will pay for a 2% better model if there is a free available one.

Also it makes all the industry standard aligning with Meta's ones.

[-]

the_hillman@reddit

Scorched earth tactic to hobble the competition.

[-]

delusional_APstudent@reddit

makes researchers build on their architecture and make finetunes of their model, which brings more attention to them, cycling this process again and again
builds goodwill
kick in the shin to openai
they’ve saved their biggest model, llama 3 405b, which i personally dont think will release open weights and we’ll have to use api
meta is a big company, their net income is 12.37 billion dollars. they can afford to let a model go or two

[-]

MicBeckie@reddit

Even if Meta releases the 405b model, I can't imagine that any of us could host it. Most people will have to rely on an API and most of them will go to Meta anyway.

[-]

TheTerrasque@reddit

You can probably make it run on a 200gb+ ram machine on CPU. It'll be pretty slow for sure, but having it run at all is pretty great.

[-]

HibikiAss@reddit

generate at 100 token per day

[-]

TheTerrasque@reddit

Some quick napkin math using the assumption that the memory speed is still the limiter, and based on the speed of llama3 70b on my server, I'd expect around 20 tokens a minute.

That's with older ddr4 memory though, but with dual xeon bringing up the number of memory channels

[-]

jart@reddit

20 tokens a minute is a lot faster than I write essays. Prompt processing speed is always more important though, and on the latest Threadrippers LLaMA 405B should in theory go ~6 tok/sec, which is also faster than most humans read. Frankly I don't mind grabbing a coffee to wait for a smarter answer. That's why I wish they would release the bigger models. It cost me about the same price of a 24gb graphics card to put 512gb of ram in my computer. Which is nice, since it's cool to be able to run models like Grok and Command-R at BF16. Although I feel a little robbed not having the mightiest LLaMA model.

[-]

Organic_Muffin280@reddit

What kind of motherboard supports this kind of ram

[-]

jart@reddit

I imagine any workstation or server motherboard. Gamers have traditionally wanted to have as little RAM as possible, so it's probably best to avoid gaming hardware if you want to run large language models.

[-]

SryUsrNameIsTaken@reddit

Local AGI is hard, okay?

[-]

shroddy@reddit

On an AMD epyc server of the latest gen and all memory slots equipped, it should be about 1 token per second.

[-]

Fusseldieb@reddit

WAW!

[-]

MoffKalast@reddit

If it were sparse it would've been doable, but alas

[-]

MicBeckie@reddit

It's not just pretty slow, it's unbearably slow! I therefore remain of the opinion that it is out of the question for most people.

[-]

Spindelhalla_xb@reddit

That’s what I don’t get about the moaning why it’s not released yet here and other places, apart from a select few no one will be running that

[-]

Expensive-Paint-9490@reddit

Fast generation is a must for many use cases, but not for all. For creative writing, editing, summarization, and many professional workloads, you just start the inference and come back when it's done.

I think Meta is not going to release the 405B model. But, whereas it would, I am eager to use it on the workstation.

[-]

Organic_Muffin280@reddit

Bingo. Then they rent you the servers that CAN host it... .

[-]

Waste-Time-6485@reddit

if we could have some sort of dynamic quant loading like using low quant to give the most prob path for the respose and then somehow load biger quant dynamically for that path and work with the alternative paths and finish the response then it would work, at some point...

anyway even with ternary weights 405b seems fat for my laptop GPU to run locally :|

[-]

gmdtrn@reddit

Yann LeCun said on X Llama400b will be open weight.

[-]

Pacyfist01@reddit

Well ... the "out of box" Llama 3 70b (without quantization) requires 140GB of VRAM.
That's already outside of what I can afford.

[-]

MicBeckie@reddit

I am in the 2xP40 club and am therefore lucky enough to be able to run Llama 3 70bQ4. I "only" get 7 tokens per second. That's just about enough for me. 405b would be extremely slow even with 8 Tesla P40 and therefore completely unusable for me.

[-]

My_Unbiased_Opinion@reddit

How slow you think?

[-]

MicBeckie@reddit

I don't know. Maybe 1 or a absolut maximum of 2 tokens per second? And presumably with electricity costs that are significantly more expensive than API prices. I'd rather go for smaller models, MOE or the API.

[-]

x54675788@reddit

without quantization

Which is why I use quantization and run it on my gaming laptop upgraded with 64GB of DDR5

[-]

Pacyfist01@reddit

If you use RAM + CPU you get much worse performance, and each level of quantization decreases model accuracy. So it's possible to run it, but not optimal. I have a PotatoPC so I'm stuck with 8b.

[-]

My_Unbiased_Opinion@reddit

P40s. Mostly us P40 folks will run it since we don't have to sell our kidneys trying to buy the cards. Should run with 4 P40s with some serious quanting.

[-]

Cool-Hornet4434@reddit

There may be future upgrades coming down the pipe that will allow us to run larger models on smaller amounts of RAM. There's another reddit post talking about it as if it's going to drive down hardware prices (I don't believe that for a second though).

https://old.reddit.com/r/LocalLLaMA/comments/1dptr6e/hardware_costs_to_drop_by_8x_after_bitnet_and/

I imagine a 1.5bpw version of a 405B model would still be several gigabytes more than I have available, but it wouldn't be impossible to run on consumer hardware.

[-]

MicBeckie@reddit

Yes, maybe in 2 years we will have the opportunity to run Llama 3 405b. But in 2 years we will probably have Llama 5...

[-]

Ylsid@reddit

They have been very clear several times over 405b will get open weighted

[-]

ahjorth@reddit

Agree with all these, and would add one point: Recruitment. Many of my PhD-colleagues who worked with machine learning in the mid 2010s wanted to go work at Facebook purely because they could work on implementing pytorch, and that would be super interesting work. (Well, and the salary.) But showing prospective researchers that this is a cool place to build long-lasting software is a huge talent magnet.

[-]

Budget-Juggernaut-68@reddit

Not just building long lasting projects. But projects that only companies like Meta/Google/Facebook can execute. Most of these models requires data centers of GPU to train for months, something working in academia or any other companies will never achieve in the short term.

[-]

ahjorth@reddit

Definitely that too!

[-]

Redhawk1230@reddit

Man I wish more companies operated on point 5. So many companies, even the wealthiest in the world, are so anal about min-maxing the profits of any project or endeavor it’s annoying

[-]

Budget-Juggernaut-68@reddit

Isn't that the whole point of a listed company no?

[-]

mileseverett@reddit

Not just the researchers building on their architecture, there's also an insane amount of tooling created for local LLMs which they can benefit from

[-]

After-Cell@reddit

There's plenty of stuff explaining open source. But what did they get in return with this actual approach?

A few:

1) recruiting. People are already familiar. Bigger recruiting pool and save money. Zuck actually talked about this. It's very competitive out there, and this made a big difference for them.

2) attention. Steisland effect. I don't need to explain the attention economy, but at the least, it puts Facebook at the centre of everyone's minds. Chatgpt has that effect from breaking ranks to be the first, I know. But chatgpt is just a product; there's not as much to talk about

There's the main ones?

[-]

I_will_delete_myself@reddit

Two words. Free research

Meta gets free testing for models on all machines, which benefits them by allowing their employees to run their models without needing a team.

The free research from normal consumers is also great because it lets people FORCE their models to work on low resource, despite the previous trend being impossible to run unless you got a army of GPUs.

Meta also benefits for it as a SDK for their VR games to create interactivity to their games.

Yann LeCun and like minded researchers also are attracted to Meta. OAI used to do the same tactic to get researchers.

[-]

FPham@reddit

Meta business model is to hook as many people as possible to use their platform and for longest time possible while the platform is seemingly free. They are not interested in selling services. Selling access to AI model like OpenAi would simply not work for META customer base. So the only way to get on bandwagon is to make it free. And the only way to save on research is to make it sorta-open-source, but not really open source.

[-]

bbu3@reddit

The following is just my opinion and what I think is plausible. I am by no means sure that this is THE reason behind their actions:

I think in the long run, both Meta (and also Apple btw) somewhat bet on giving AI capabilities for users for free but then possibly make money from what users spend through AI.

E.g. if your AI workflow results in products and/or services being bought, there will be money for the maker's of AI models be made from the businesses that offer these payed services. Meta may bet on being a provider of such an AI model, powered by all the user-specific context that get from their networks. Meanwhile, Apple doesn't need their own AI model. Theyt can make deals for which model provider to make the default for their wealthy userbase.

Thus, it would be similar to search, where users don't pay but Google gets money from the businesses that want to sell stuff on the web and need Google Adds/Search to bring users their way. Apple, in turn, gets (or used to get) their big share from Google just for pointing IOS users their way.

If all LLMs are nearly equally good, contextual data around the users makes or breaks the AI service value to users. In that case, Google might be serious competition to Meta, but for everyone else it's hard to catch up them in terms of knowing the users. Thus, it is in their interest if the models themselves are not the distinguishing factor.

I think the big goal for Meta would be that users get to use AI models for free, but companies pay to "advertise" / cooperate through the AI responses. Meanwhile OpenAI (and Anthropic) try to have the very best model so that users pay them to use the model.

[-]

Illustrious_Cook704@reddit

Apple is betting on offering AI for free ? free, but only on their overpriced products. They have never offered anything...

[-]

bbu3@reddit

My point is, rather than build their own foundation models and compete, they hope AI turns out like search: Apple doesn't need to compete, it sells their user base. That works better if they sell to a "free" service, because when users are paying directly, they want to make the choice for themselves

[-]

Illustrious_Cook704@reddit

The behavior of Apple is really becoming more like extortion than anything else, they are obsessed with money and nothing else matters really. They keep increasing prices, and are very conservative, not innovative. Their adepts and fans are believing anything, but it's a reality that not only no competitors are allowed in any way anywhere, this is the most closed OS that exists, you don't own the device, nor the OS, and can't even make decision on so many topics. This is

[-]

ChrisAlbertson@reddit

They did not actually give away $100 million because they still have it. A better way to say it is that Meta is letting people use their $100M product for free.

When you say it that way the economics is clear, lowering the price of any product increases the sales volume. Meta wants market share and how better to get it than by lowering the price to "free". Market share is what makes your work relevant.

Also and maybe more importantly, maybe Meta is seeing that the sellable product is not the model but the software you wrap around the model. They have nothing to lose by giving it away and some to gain.

A final thought is that these LLMs are like bananas. They have a short shelf life. So while technically they gave away Llama3 "forever", it will have lost its value in a couple of years.

[-]

mikebrave@reddit

Short answer is they feel somewhat threatened by google and openAI, they can almost compete with them but not completly, so to keep them in check they released their model for free and allowed the community to get involved thus improving on what they already made. Which is basically actually keeping them somewhat in check.

[-]

aifreakshow@reddit

Good question

[-]

ineedlesssleep@reddit

This post reads so much like a ChatGPT generated response.

[-]

DominoChessMaster@reddit

Gemma 2 is better and less than half the size. This means you can use it free on your computer and it’s going to cost your business way less to serve requests to customers.

[-]

x54675788@reddit

I am willing to bet they'll stop giving models for free the moment they are actually better than any competition.

No, I don't think it's a "good will" move. The moment a model stronger than the GPT4o (and all the others "Pro" cloud models available now) comes out, I am quite sure it'll be too "unsafe" to release to the general public for offline use, but (yes, you guessed right) it'll be available online only through their platforms.

[-]

syrigamy@reddit

The you don’t know Meta, they’ve been giving technology for free for a long time.

[-]

Illustrious_Cook704@reddit

Indeed, for AI, xformers and segment-anything are examples (even if SAM has a lot of "concurrent"). React, original torch. They also publish scientific papers, like Microsoft who published thousands or papers (I use this example because when I studied, the cloud was taking off, and there was a need for new network protocols and scaling techniques, etc. and it was astonishing that they provided advanced tech with experiments etc. for free, when some companies barely share anything...).
So some interesting software, but nothing so revolutionary... Also, now that I think of it, Meta us a bit shady company, overall, why do they offer all this ? Maybe to also hqve the community improve it ?

[-]

syrigamy@reddit

I’m not defending meta by any means as Meta is just an other greedy company. But I don’t like how we criticized Meta while supporting other companies that are worse. Meta probably giving all these techs for free. First because Zuck really believes in open source projects, another reason could be recruiting scientist. Most scientists work on open source projects and use them, so if they want to go for a company after getting a PhD is way easier to go to a company where the main tech you used is developed by them. Then there are a lot of scientists that directly support open source projects and only by that they’ll join a company. Some folks don’t care about money even though Meta throws few millions on their best scientist

[-]

Illustrious_Cook704@reddit

Computer science is indeed science, and as such come companies publish in collaboration with universities, or fund research. And this is a really good thing. CS is a really complex domain, and funding fundamental research, or designing and experimenting solutions for problems that are are real challenges, while making it public, if the paper gains some tractions, code is usually provided. Topics like networks (which are like 'alive', they really need to be studied in real conditions to have a complete view of the impact of a new protocol, etc.), I saw a few Natural Language Processing ones, super interesting.... the 'sad' or discouraging thing is most papers that are results of years of work, don't lead to anything. Like new RFCs, network standards, are already difficult to even be considered, but becoming an actual RFC is rare... (which is normal, networks are sensitive, and small changes can have important consequences).

It's true Meta is quite honest on their ads and data collection. Which is how they make money, 121/134B in 2023... and hey have platforms where they can see by what people are interested etc. Google, who also has a few famous open-source projects, usually presents free services as generosity, but never talk about how this will benefit them... In reality, I dont care that Google knows I looked for a tshirt on Amazon... but small business, if they want to be on the first page, spend lots of money on ads, and this is not healthy (it's a well known consequence)... moreover, now that legislation is getting more constraining, 3rd party cookies are already deactivated in Edge and Safari I think, they'll have to do use more passive solution, like fingerprinting which is highly efficient... anyway...
But when it comes to removing massive amounts of money from the economy, and pretty much not giving anything back, nothing beats the fruit company, who additionally is rarely factual on any topic.

Then, if Meta open source React, which is good for the web because it's very good, and at the same time benefit from the participation of the community... it's not evil... but I wonder if this is part of the strategy or not.... :)

[-]

x54675788@reddit

But AI (ok, ok, LLMs) are something unprecedented, not to mention regulated, and something that moves billions of dollars, military interest and talks about power plants.

[-]

_anotherRandomGuy@reddit

TL;DR- GPT is the main thing for OpenAI. Llama is not the main thing for Meta.

There is barely any incentive for OAI to release their model weights. Their (new found) revenue model effectively relies on their model inference (chatgpt plus subscription + APIs). Releasing OAI model weights let's their biggest competitors (which are Azure/GCP/AWS, and not other AI labs) in on their pie.

On the other hand, Meta has various revenue sources (ad money, etc) that keeps their servers running. As a result they can afford to spend money on R&D and open source their results for "free". Yann Lecun has spoken about this before.

Also, Zuck talked on Dwarkesh's pod about how they benefit from OSS technological innovation on top of their model releases. eg. Llama was crucial for initial model quantization research. 5% increment in inference speed for Meta translates to millions of dollars of compute expenditure savings on their various services. That's a pretty good ROI for a relative small R&D release.

[-]

vuongagiflow@reddit

They are playing longterm game and have capacity to do so. Many tools and cloud providers support llama while gpt and claude has lesser support coverage. Which means they have better bargaining power later on if they want to leverage that.

[-]

Hyp3rSoniX@reddit

Maybe they're able to convince more investors this way that they indeed are capable of doing AI.

They did hurt their image before with the whole Metaverse thing after all...

[-]

AnxiousWolf9176@reddit

How does open source model ensures low competition for closed source open ai? Like does your average person cares? (A genuine question because I don't know)

Also does being open source means they provide everything to run them model locally free of cost, or you still have to send requests to meta to retrieve the data and it just doesn't costs anything?

[-]

Illustrious_Cook704@reddit

The model has a quite limited licence and isn't open source. It's not really close to GPT-4o, because it isn't multimodal... but the community added vision to it...
I's a good model, but there are others good models built by smaller companies too.

It's not entirely evil or wrong, but I believe one of the main goals is having the community improving and creating tools for free... good tools by the way.

But Meta, who has barely any paid product (I don't even think the marketplace charges fees), who is selling data to thousands of companies, which isn't a secret, but still are a shady company suddenly becomes so generous... It's hard to believe they are completely honest...

Then about the OpenAI model... the few companies that have very good models and make money... have closed models. The only exception is Mistral, but it's not sure they're profitable...
OpenAI has spent billions and has made years of research to reach that level. They publish scientific papers, but they have to pay the bills... they still offer free services. Same for Anthropic...

Being a company that has created cutting edge technologies, it's normal that they want to keep it closed.
A company like Stability AI has created models that are widely used and also were a revolution, but they had barely any source of income... and they almost went bankrupt. It's not possible to offer everything for free and then spending huge amounts of money to do it... Being a company and selling products, is normal...

Microsoft is indeed offering a lot for free and I think that this pushed Google to also offer something for free. If Microsoft had decided to make it paid only.. Google would probably never had done it either, same for other companies. I think this had an impact on the economic models...

[-]

Organic_Muffin280@reddit

Are there more open projects with similar performance to lama?

[-]

Illustrious_Cook704@reddit

I try to follow the news on AI tech etc. but there are so much happening everyday it's impossible to keep up.

This is what people designing model do: new strategies are imagined by computer scientists or companies, then model designers apply various of those strategies, to optimize models, in different ways with different goals... Making models cheaper to train, creating smaller models with matching the quality of larger ones, a practical case : some companies implement mixture of experts models, which are composed of smaller models specialized in particular tasks, instead of one big model... models which replies follows a strict formal structure, some models are optimized for coding, others for robotics purposes. or finding new ways of structuring the model file itself, optimizing or creating new algorithm to improve memory usage, speed... another example: mamba state-spaces/mamba: Mamba SSM architecture (github.com) I don't know the details at all but it's apparently a big evolution about the way current models operates...

There are lots of more open models, or probably also many more restricted ones. But I can't give any details, "performance" is not a general property, you can be performant in one area, and not that good in another... be faster, be smaller, be more accurate, better at math, be more efficient at memory usage, being cheaper to train...

Specialists may have more informed opinions, but I think there isn't ONE model that beats every other in all ways... GPT4 has been quite effective at being at the top in benchmarks (benchmarks are not really perfect and an absolute ways to rank models, sometimes even not neutral or honest) ...

[-]

h4xz13@reddit

If everyone has it, then nobody has a upper hand

[-]

martinkou@reddit

Commoditize your complement - it's the oldest trick in the Silicon Valley business playbook.

Meta has a lot of proprietary data and consumer / business entry points and platforms like Instagram and Facebook. The last thing Meta wants is some AI up-start with a groundbreaking model being able to surpass them in user experience.

By giving out large models, it flattens the market on the model-building dimension - i.e. the business value of any model research alone when used in consumer use cases, is close to $0, no matter how expensive it cost to build.

It also means the complements to advanced models become even more important and valuable than before - i.e. any progress on LLM models would only make proprietary data and proprietary platforms more useful, and harder to beat.

[-]

Zulfiqaar@reddit

I've moved over a third of our usage to LLaMA 3 and Command-R+ (and soon Gemma-2) - each of them are better than last years GPT-4. Theres a threshold of "good enough" for a usecase, and then its just a matter of speed and price and convenience.

[-]

Organic_Muffin280@reddit

Do you believe is the reduced censorship that makes it superior? Because frankly the mainstream models are getting out of hand. Even the tiniest inconvenience makes it a woke machine that "cannot elaborate in insensitive topics"...'.

[-]

Zulfiqaar@reddit

Smarter? Honestly not really, as this year's iterations by OpenAI and Anthropic are smarter.

It's more that the new Open models are smarter for certain use cases than proprietary from before, and they have a nice generation speed. Price reductions by 20-60x is nice too.

And lastly for one usecase (browsing assistant through ChatGPTBox) the uncensored models are far more useful, as that's where I'm actually likely to encounter a input context that is likely to trigger safety features that Llama or CommandR won't have any problem explaining or summarising

[-]

redditrasberry@reddit

i think people grossly misunderstand Zuckerberg

why is he doing all this? Because he spent all his formative years in tech struggling against other platform owners who controlled his fate completely. He could do literally nothing if Apple, Google, MS put him out of business.

What is a natural reaction to that? Once you have the power, you swear, you will not be in that position again. How do you do that? You do everything in your power to control the platforms you depend on. The realistic way to accomplish that is to control the platforms the whole ecosystem uses. How do you do that? By building them and distributing them for free - best in class, open source, get everybody using your platforms. Then you are the one in control.

This explains Llama, it explains PyTorch, it explains React, it explains his obsession with owning VR. You just have to understand this one thing about his life experience and everything else makes sense.

[-]

Organic_Muffin280@reddit

Nothing is really for free or open source in the minds of billionaires. It's all a strategy and a reaction to their market pressures

[-]

sobamf@reddit

When you’re ahead, closed source. When you’re behind Open source.

[-]

Organic_Muffin280@reddit

And if you gather enough simps, can get closed source at your next "premium" product (that no home PC/server can run anyway).

[-]

KallistiTMP@reddit

For starters, the strongest supporting software ecosystem for any LLM on the planet, hands down, at a development cost of $0. You want it quantized, on a phone, with LoRa, with function calling and four types of infinite context? Llama's got it within a week of the first research paper.

Llama is not the product. Llama is just the tooling that Meta uses to make their products. They get an absolutely massive amount of totally 100% free volunteer labor to improve every aspect of their toolchain. They have probably saved tens if not hundreds of millions of dollars on reduced hardware costs alone just due to community built optimizations.

Also helps attract talent, reduce training costs, etc. In a market where skilled AI researchers are making million+ salaries, it really helps to be the cool kids on the block that everyone wants to work for, and for your new employees to already be intimately familiar with your product and tooling from day one.

This is of course only possible because they have directly monetizable applied use cases for LLM's, which are vastly more profitable than selling LLM inference services. If they can sell 5% more ads thanks to community improvements to their Llama models, that 5% bump is likely bigger than OpenAI's entire revenue. Why bother selling shovels to prospectors for $5 a pop when you're sitting on the world's largest gold mine?

[-]

Mission_Tip4316@reddit

Could you share more on how do to LoRa for function calling on Llama?

[-]

KallistiTMP@reddit

At the same time? The simplest way would be to train a lora on your dataset (presumably a dataset of function calling examples suited to your use case) using unsloth, hugginggace PEFT, or any other readily available tooling for LoRA or qLoRA. Then merge the adapter into your weights, load it into your function calling framework (llama.cpp, guidance, etc) and you're off to the races.

I believe the huggingface TRL library also supports reinforcement learning with function calling directly, which may be more suitable if you have a use case where your function calling translates well to a reward model.

[-]

Mission_Tip4316@reddit

This is what I think would help me reduce the cost of my function calling chatbot based SaaS. Currently all SOTA models are very expensive specially with chat history(context).

I would really appreciate if you could guide me towards certain keywords or search terms I can use to read more, best if I can find some YouTube video.

Just a quick question, what are the chances that with a good training set ( based on Gemini 1.5 or Claude 3.5 based dataset) I can achieve 100% accuracy with enough samples on let's say Llama? Because I have tried function calling specific LLMs and the problem usually with them is a small input window.

[-]

KallistiTMP@reddit

Just a quick question, what are the chances that with a good training set ( based on Gemini 1.5 or Claude 3.5 based dataset) I can achieve 100% accuracy with enough samples on let's say Llama? Because I have tried function calling specific LLMs and the problem usually with them is a small input window.

It's impossible to say without knowing the use case. Accuracy in terms of making a valid output schema is guaranteed for native function calling frameworks (llama.cpp, guidance, etc), which use logit bias and grammars to ensure that the model can only select valid tokens.

Accuracy in terms of calling the right tool at the right time with the appropriate arguments, that is 100% contextual. If you're trying to make it recognize if a URL is valid, then sure, 100% success should be doable. If you're trying to make it pick winning chess moves or accurately screen for different types of cancer, that will probably be a lot harder.

I will say the biggest recommendation I can give is don't fine tune. At least not until you're already getting good results. Old school ML people usually reach for the fine tune knob immediately on instinct, because that's how old school models worked.

LLM's are fundamentally different, and far more sensitive to how the input data is structured and what prompt engineering techniques you're using. Prompt engineer first. The vast, vast majority of use cases do not benefit from fine tuning, fine tuning will most often just degrade performance in a way that overfits your training dataset.

It also has a widely observed overfitting effect that hasn't been named yet as far as I know, which I refer to it as the "precision bullshit generator" effect. Essentially, training new forms of deep reasoning is hard, but training surface imitation is easy. So what can happen is you can end up training your model to generate output that looks plausible at first glance, but is actually complete ungrounded bullshit that just happens to use the right buzzwords and format to make it look convincing. This precision bullshit is often much more harmful than obvious bullshit. The Curl team has a good write up on how this has affected them directly.

That said, yes, with good prompt engineering most open models are actually very capable.

[-]

Organic_Muffin280@reddit

How will they cheaply expand the context

[-]

Mission_Tip4316@reddit

I have had something similar happen when I tried to fine tune the gpt-3.5 model and it ended up outputting only function calls and lost all of its other general abilities, maybe my dataset was also not the best back then.

But the primary reason I am trying to do this is because my use case uses an API to get details about holiday packages which contains a lot of text, so maintaining context is very expensive with sota models such as gpt-4o and gemini-1.5-pro. I have been perfecting the prompt but it ends up being close to 5k token just the system prompt.

To be able to sell this as a solution I need to control costs and that's where I am lost and there only a maximum that I can remove from the holidays api response, on the other end using cheaper models has resulted in a lot of mix ups between a city, country, continent etc names when using the tools or not always using the budget range parameters of my functions,

So I have been trying to learn if fine tuning or creating my own custom model using some open source model with atleast 32k context window if not Higher.

If everything fails I might have to offer a bot that forgets context every few prompts ( I have tried using summary etc but it ends up adding latency issues)

[-]

KallistiTMP@reddit

But the primary reason I am trying to do this is because my use case uses an API to get details about holiday packages which contains a lot of text

How much of that text is useful? How much can be filtered or fetched dynamically piecemeal through an agentic tool calling approach rather than dumped en masse?

You're not limited to a single tool that dumps the whole API response into the prompt, splitting up into smaller subtools might get you where you need to be. I.e. even if the external API returns a giant monolithic JSON object, you can still create tools that only return specific fields. Break up get_package_info(package ID) into get_package_location(), get_package_sender(), etc.

If it's actually an unstructured text extraction problem (i.e. the bulk of the useful data contained in the API response is messily scattered throughout a free-form "description" string field) then you can consider using a much smaller and faster model for extraction. There are actually some very nice models specifically fine tuned for translating unstructured text to a defined target JSON schema, on the order of 7B parameters or less, so lightning fast compared to your big model.

If your context is large because of a large number of few-shot examples (which, if you aren't using few shot examples, you absolutely should be), you can safely roll that out of the context window without consequence in many cases. The initial generated content can act as examples for subsequent generated content, within reasonable bounds.

Just throwing out general knowledge here though, if you want to share some examples excerpts that illustrate why your context is growing so large I can probably give you more directed pointers.

[-]

Mission_Tip4316@reddit

These are some very good pointers thank you, so just a little background: I am your chatgpt programmer but have some ability to read and understand code.

Tools: So I started with my API provider's response a few months ago but since then I have pivoted to using Elasticsearch indexes for me to create better search functionalities and my own query parameters (so I am not limited to the filters they are providing)

I am still learning how to effectively use Elasticsearch and it's ML capabilities to query documents.

But yes, I have now created my own abstracted API and tools on top of it for the model to use and query the holiday packages based on vague user queries and providing results with minimum inputs from the user (that's the USP I am trying to build) currently the Gemini Model accurately handles 5 tools with some nesting for parameters, I am yet to experiment with more nesting of parameters to get more filtered data while the user has questions about several aspects of the holiday package

Regarding Context: my context or chat history is growing large because of tool's responses mostly - One tool's response contains for example prices: a holiday package can have 30 different kind of deals available which when I tried to nest results in a lot of errors by models

So the solution I came up with is to implement UI to show the information from tool's response and send minimum required information/context back to AI model or in the chat history to minimise the cost.

I'd definitely check out this advice to remove example after a few prompts

[-]

EmilPi@reddit

I think that the opinion is correct, that Meta releases its models to make some users not pay for OpenAI's models. But general users win from this.

[-]

cloudkiss@reddit

Good intention. I appreciate Meta’s decision.

[-]

keepthepace@reddit

"Look, over the next 10 years, we are going to spend about 100 billions on AI inference. If the open source community can make this process 10% more efficient, then that's worth spending 10 billions on it"

Mark Zuckerberg.

[-]

Kanthumerussell@reddit

Yeah this is exactly my thought. In a sense I think OpenAI is playing a very risky game. I mean I felt like we saw this in what seemed like the overconfidence of Sam Altman saying there's no way anyone is going to be able to touch them in terms of how advanced and capable their model is. But thats assuming the infrastructure is in place and now its just a matter of who can compute more. But as we've seen there are constant improvements, some pretty major in the development of these models. At an given point in time someone could come out with something where it basically means a model as capable as GPT-4 could be built with a fraction of the resources it took for them to build it. This of course is always expected to some degree in this industry but I think right now its especially volatile and I personally don't want to be the dude that spends all of my money on what could cost pennies next week.

[-]

jsdfljsd@reddit

Meta is known for contributing many items in open source like GraphQL, React, and so on. So, they might want to continuing the effort in open source

[-]

arcane_paradox_ai@reddit

META stock was doing really bad when tey were in the hype of the Metaverse, triple the stock price when they got into AI. They were late they can rule Open Source, that will attract talent. And many companies will start using it, because most large companies don't want to send their sensitive data outside their datacenters, these companies will get used to Llama and will be easy clients for paid services from Meta, support, consultancy services, or more advanced paid models that might run in local datacenters, the step number one is to get them using your product even if it is free! How many Open Source products does Google or Microsoft have? Loads of them!

[-]

P8ri0t@reddit

Freemium model..

They give out a limited version for free that people like using and will adopt into their daily routine.

Then they show the advantages of their paid model and offer subscription services.

..like a drug dealer offering free samples, but more effective because everyone's doing it.

[-]

ron_krugman@reddit

What exactly locks someone using Llama 3 8b or 70b locally into Meta's ecosystem? It's just an LLM like every other model on the market.

The whole point of open-weight LLMs is that I can run them locally. If I can't, there's no reason to prefer Meta's products over those from OpenAI, Anthropic, Google, etc. because at the end of the day they're all just prompt-response machines.

[-]

P8ri0t@reddit

If I understand, you're saying that the selling point of running their latest model in the cloud wouldn't benefit you as much as simply using their local models or those of their competitors?

I agree. Knowing about the industry, you would benefit most by simply using the best free locally running model available rather than beta testing the latest version at the expense of your personal data.

[-]

Best-Apartment1472@reddit

It's not "free". It is just binary file. No way you can re-train it just from that. Even if you can, nobody have compute for that, at least nobody in small, medium size company. Also it could have build in biases that are aligned with Meta's values, which is by itself very powerful.

[-]

gmdtrn@reddit

The weights are there specifically for fine-tuning. So yes, you can and many already have trained from that. While not inexpensive, a high end consumer PC can indeed fine tune (slowly) with even the 70B model.

[-]

Best-Apartment1472@reddit

Fine tune sure, re-train model to improve and build stuff upon it. To learn how to build such model. That is not available. You can also fine tune closed models...

[-]

gmdtrn@reddit

There are quite a few great tutorials on building LM's. Of course, there's no doubt that there are some proprietary secrets. But, Andrej Karpathy literally coded GPT 2.5 on uploaded it to YouTube. The major takeaway is that, largely speaking, the compute is a larger limitation than the code for pretty much anybody who isn't a MAANG company or funded by a MAANG company.

[-]

HarambeTenSei@reddit

Meta's business model is advertising and abusing user data. It's not AI.
By releasing these AI models they bribe people into using pytorch, advertise themselves to potential hires and most importantly: reduce the potential revenue stream of openAI/microsoft/etc by creating competition for them through open source.

[-]

Ousseraune@reddit

Tell me. Is Linux or Windows better?

Let's say if using Linux, they had someone to customise it to your needs.

What wins.

Llama is amazing. Maybe not as good as the best GPT atm. But open source is their bread and butter and helping them improve their software. Yes, at some point they'd need to make some money from it. But until they feel ready to dethrone the king, having the people behind you will always be a huge boon.

[-]

BlueShadowsOfLight@reddit

..the real reason is, now there is no legality for zuck to sit in front of senate alone.. cant be charged if so many are doing the same

[-]

phhusson@reddit

My 2 cents:

It is very possible that the reason they released it is simply that their researchers wanted to. Publishing it doesn't harm Meta's business at all, and you keep your researchers happy (also you can hire a bit more smart people)
Beyond that, Apple/iOS and Google/Android walled gardens are big dangers to Meta: At every new version, Apple/Google locks users in more and more, which means that Apple/Google gobbles up more and more user data (I mean it as "only Apple services can use those data", not "Apple can read those data"). Which means that overall users spend less and less time on non-Apple/Google properties (where Meta make money on most websites).

Giving a free model to the general public means that smaller website/services can better compete against Google. And a huge proportion of those "smaller websites" will use Facebook services, thus giving more money to Facebook.

[-]

sebramirez4@reddit

I think meta thinks more about AI as a tool and a feature rather than a product in and of itself, by open sourcing it they get people running it with fewer and fewer resources which helps them optimize the way they run their own models for their AI features, I personally think it’s the most rational and best approach because honestly intelligence in an AI model should be less important to how useful it is and being able to run your models with fewer and fewer resources than your competitors is really valuable for building out good and fast features, it could be a wrong approach because of how fats GPT-4o has gotten but for a company that started way behind openAI I think it’s been amazing for them.

[-]

Ok-Excuse-4371@reddit

might want to do a reddit & google search before posting.
many posts & articles about this already:

[-]

Worth-Card9034@reddit (OP)

Agreed with you. I am little naive with using Reddit, since i saw ML folks are hanging here. I did put my question in search bar on Reddit, however it didn't list my question. I am definitely not good as the way you searched. I will keep that i mind in future!

[-]

East_Professional_39@reddit

Use perplexity, before searching click on the Focus button and choose Reddit, it's pretty good

[-]

Ok-Excuse-4371@reddit

Thanks for reminding me of this.
After hearing about it for months.
You got me to finally try it!

[-]

trotfox_@reddit

Ok I came back to say, holy shit!

This is great stuff, I cannot justify 20 bucks yet, BUUUT, if it ends up blending itself into my day as my new search/assistant/????, it just might be!

Can anyone who has pro use the "create images" of your search feature, I need to know what that does!

[-]

trotfox_@reddit

Me too, gonna check it out finally

[-]

Downtown-Case-1755@reddit

Random tip, use old.reddit.com

[-]

Ok-Excuse-4371@reddit

You are very gracious.
Reddit search is kind of weird because it doesn't like spaces.
Luckily, Google searches Reddit better than Reddit ;-)

[-]

ARush1007@reddit

Reddit search is pretty trash. I find much better posts related to the same queries using Google. Duckduckgo is sometimes even better for certain topics, subtle bias/censorship is crazy popular these days.

[-]

HyperFoci@reddit

Give away the hobbyist models <120b or so, and promise the >400b ones will be even better but will cost ya.

[-]

ROGER_CHOCS@reddit

100 Million might be a lot for us, but it's nothing for a company like Meta.

[-]

mythicinfinity@reddit

The future will be dominated by AI companies, and one of the largest points of competition for all tech companies is over talent. They have to either compete in the space or lose out on the future ability to competitively pay talent.

[-]

BGFlyingToaster@reddit

I know it feels like the GenAI industry is all about who has the best model; as consumers, that's what feels most important right now and we'd be correct in thinking that it's very important to have good models. But these companies aren't prioritizing model development; they're prioritizing the building of compute capacity. Meta, Google, Microsoft (with OpenAI), and Anthropic are all building massive ($100B+) compute centers (aka data centers) so they can handle the world's quickly growing needs for AI. Having lots of good models out there benefits all of them. It's why Microsoft gave away Phi-3, why Meta (in partnership with Microsoft on V2) gave away Llama 3, and why they're all investing in hardware in one form or another. Pretty soon, good models will be a dime-a-dozen and all that will matter is who has a compute to train them, fine-tune them, and most importantly - run them.

I work in tech consulting and my clients are already feeling the squeeze for compute. We're being asked almost daily to spin up massive workloads to process information with GenAI - every email, every document, every business transaction - billions and trillions of things that need to be analyzed, summarized, categorized, and processed.

[-]

MrOaiki@reddit

Because Meta makes money for their data and user base by selling ads. AI companies like OpenAI make money off language models. If you can release a good enough language model freely, you basically eradicate AI companies and stop them from ever getting to a point where they start competing with you.

[-]

farhan3_3@reddit

It devalues OpenAI and GPT-4 by giving something competitive for free.

[-]

Illustrious_Matter_8@reddit

As opensource they allow for improvement standards and other people improving upon it. We've seen the effects of that with other neural networks stable diffusion got Lara invented and lots of other improvements. We're now at the start of artificial intelligence, yes this is just the start. So give it another 20 years and then you l notice that spin things up you better seed early. So who will benefit long-term of this Meta

[-]

RiffyDivine2@reddit

Give the tool for free to try and get it to be the standard everyone uses and then charge for it.

[-]

Passloc@reddit

But the thing is the models are improving so fast Llama 3 will be obsolete in 3 months.

[-]

Ylsid@reddit

This question gets asked /a lot/ here. Even meta themselves explained why. In short, they aren't an AI company, contributing to OS is core to Meta, and everyone will end up using llama

[-]

ACuriousBidet@reddit

It's not "for free" if and when it starts making a lot of money.

https://llamaimodel.com/commercial-use/

[-]

mobinsir@reddit

Torched earth strategy basically, if they can offer for free and have on par performance than others that’s paid, will people more likely choose free or paid

[-]

mmmm_frietjes@reddit

Didn’t training it cost 10 billion? 100 million seems very low

[-]

gmdtrn@reddit

You are correct. Not sure why you got downvoted. It’s a $10bn model and the 400b+ model is still in training.

[-]

mcharytoniuk@reddit

Most people won’t make an extensive use of it and i think it doesn’t pay off for them to give support for small businesses. They still charge you for it if you have enough revenue or users.

[-]

gmdtrn@reddit

Llama3 70b is like 20x less expensive to run than GPT. 8b even less expensive than that. GPT is overkill for many inference tasks. Huge volumes of people will use Llama commercially.

[-]

JockY@reddit

I could be persuaded that Zuckerberg actually believes it's the right thing to do.

[-]

MoffKalast@reddit

I don't think anyone will try to persuade you of that.

[-]

JockY@reddit

There is no evidence to suggest he believes it’s the wrong thing to do, and $100m of open weights to suggest the contrary. He showed us who he is, I say we believe him.

[-]

ahjorth@reddit

It is of course open to interpretation, but he has shown us who he is since the early 00's, and to me that has provided no evidence that he does anything because he thinks it is the right thing to do.

[-]

gmdtrn@reddit

Nothing except open source a giant set of high value tooling that’s nearly revolutionized engineering. Well before Llama3 Meta has been open sourcing the best tools and frameworks in the industry. Many of the worlds oriente web apps are run on frameworks written by meta, and basically all of the high value ML solutions written are done so using libraries written by Meta. Llama3 is just the next tool in a like of many they’re already doing this with.

[-]

JockY@reddit

You’re presenting a lack of evidence as evidence itself. Good luck with that.

[-]

ahjorth@reddit

It was meant as a direct response to your comment that

He showed us who he is, I say we believe him.

If you see no evidence that he has consistently ignored what was the right thing to do in favor of increasing marketshare and killing competition, then we see the world very differently.

[-]

dwaynelovesbridge@reddit

This is my controversial opinion as well. If you listen to him on Joe Rogan, it’s pretty obvious to me that he is really still a nerd at heart who just happens to be the founder of what has become an extremely evil company, but he himself is not a bad guy.

[-]

johnnyXcrane@reddit

Of course he believes its the right thing to do or he wouldn’t do it. But OP asks for the reason why they think its the right thing to do.

[-]

bigmanbananas@reddit

Lol. I love the humour here sometimes.

[-]

cddelgado@reddit

Meta gets a few things out of it.

Free testing and updates - The Llama 3 "brand" will be tied to mostly positive things and there will be buckets of mad lads like this motley crew on Reddit willing to bend, break, twist, pull, push and everything else that could potentially make it better or more capable--for free. They don't need to invest in the GPU hours when an entire community will at no charge?
Risk mitigation - The first Llama was leaked. When you release it in the way you want, you control the situation and manage the risk.
Community good will - this is a PR win for Meta when it comes to potential talent.
Ubiquity - IBM-compatible PCs led the market for years because IBM was forced to accept that companies could make their own reverse-engineered products (I'm looking at you, Compaq). Llama, Llama 2 and now Llama 3 are now the thought foundation for so many others and Meta can claim ownership of that because they were the open noes, not the people who said you can't touch it.

[-]

gwicksted@reddit

They probably couldn’t easily monetize it themselves so chose to simultaneously hurt OpenAI and help the open source community while also showing themselves as an AI powerhouse. $100M isn’t pocket change but it’s also not game changing for them.

[-]

Morphon@reddit

Because, like Amazon (which uses AI extensively for things like summarizing thousands of comments), Meta isn't selling AI as a product. They're using it as part of their infrastructure. As such, they benefit dramatically from all the research and innovations that are built on top of it (or that used it as the best openly available model out there).

Why spend all the time finding all the possible optimizations and use cases when they can let everyone else do that and reap the rewards for free?

Again, they're not selling LLMs. They're using them. And they'd rather use their own than pay a company like OpenAI per token.

[-]

soBouncy@reddit

My wild guess:

It was trained by crawling the open internet using whatever it could find, regardless of copyright, privacy, accuracy, etc. Perhaps by releasing it publicly they might somehow relieve themselves of some legal liabilities for scooping all that data, as opposed to turning it directly into a revenue generating product

[-]

am2549@reddit

My theory:

We are entering a stage of humanity where AI (and maybe AGI) is the means of production. AI is a combination of software and hardware. In the end the means of production are the GPU farms, the digital oil. Meta, Microsoft, Google and Meta are all ramping up immensely. Meta was lagging years behind, so the combined approach of selling software and hardware is not working. So their bet is to level the playing field my taking the software component out of AI (making it accessible to everyone) and only make it about a hardware race. And this is something where Meta could succeed.

[-]

magicalne@reddit

Nobody mentioned in lex's podcast with Yann Lecun: https://www.youtube.com/watch?v=5t1vTLU7s40&t=829s&ab_channel=LexFridman
Yann explains the business model of open-source LLMs. There are many business partners at Meta. Those companies choose Llama 3 for a reason. The data is biased, and so is the model. Therefore, the only way to fight against bias is to open-source the model and let users fine-tune their own versions.

[-]

gmdtrn@reddit

It’s like a $10 billion dollar model. And Zuckerberg openly states his reasons in interviews. The short of it is that he does believe it’s the right thing from a humanistic perspective AND they believe that it makes business sense as well. Meta has open sources ungodly volumes of high value data and it’s driven their own costs down as a result.

[-]

LucyFerAdvocate@reddit

The licence means it and it's derivatives can only be used by Meta and small companies they don't care about for commercial reasons, they get a load of 'free' labour from the open source community which they can use to get a competitive advantage over their competitors. And the business models of smaller companies might become reliant on their tech.

[-]

DustinBrett@reddit

Just because it cost $100M to train doesn't mean it was worth that much.

[-]

Additional_Test_758@reddit

Meta had a lot of goodwill to makeup back from the days they allowed companies (Cambridge Analytica et al) to farm every aspect of your FB account, I suppose.

[-]

jart@reddit

It's the Alfred Nobel strategy. The greatest monsters from the ruling class become history's most celebrated heroes when they use the money they've squeezed out of people to stimulate scientific and social development.

[-]

qwpajrty@reddit

Meta didn't train this model just to give it for "free" to the community. They trained to use it for their products internally and they don't have anything to lose by open sourcing it, they actually gain in indirect ways like the other people already mentioned.

[-]

kmp11@reddit

Meta probably recognized that Claude and OpenAI will end up serving institutions and government. They will be too expensive to run robotics or kitchen appliances, gadgets, type of application. They want those tinkerers to use Meta as the default. it is not a fluke that the 70B can run fairly easily on consumer grade hardware.

I am sure that meta will get into specialized model for many of those developers which will generate licensing fees. a 70B model will be hard to fit in a robot, but a robot probably does not need 70B. It needs 8B of highly focused robotic training that can be run in 10GB of memory and a relatively simple processor. maybe model could be swapped depending if you want the robot to clean the kitchen or be a nanny.

[-]

tgredditfc@reddit

Because many people had asked this same question and got answered.

[-]

randomrealname@reddit

To absorb the open source communities creative efforts. Think of all the dev's who would actually make a good contribution that cant get a job at an ai company. The idea is if they are working with only Meta models, then Meta will be the one to gain an advantage over the other companies who are completely closed.

The hive mind of society can, and does, produce work that even professional labs cannot, this is a real phenomenon. It has happened in other disciplines like Math.

[-]

CellistAvailable3625@reddit

Because they wouldn't be able to sell it

[-]

Omnic19@reddit

they get quite a lot of goodwill from this. if it was a strategic decision whether that goodwill is worth $100 mil is a question for another day.

but where did this $100 mil figure come from. if it is the cost of buying the h100 gpus well then they will be reused again for training another model bringing the average cost down per model.

but if llama3 400b is released, Openai will most probably opensource. gpt 3.5

[-]

shikaharu_ukutsuki@reddit

:) the importance of Ai is data. Model is just the result, and be outdate soon :()

[-]

Adventurous_Train_91@reddit

Zuck has said part of the reason is that it makes the models cheaper to develop over the long term. So if a lot of people build with llama, there will be innovations and cost reductions in the software and infrastructure as well which will save them billions over the long term.

[-]

rishiarora@reddit

There is no moat in AI models there are good enough models publicly available already. The move will help to dry out the less cash rich component and will also acclimate the new audience to AI. Basically Facebook is making customers and killing competition in one move

[-]

Infamous-Scallion363@reddit

Unique things that I understand is:

Selling access to advanced versions of Llama 3, including a 405B parameter model, through an API or licensing deals.
Increasing user engagement on Meta's social media and messaging platforms, leading to more ad revenue.
Attracting top AI researchers and developers, leading to valuable partnerships and future revenue streams.
Gaining market dominance by offering a high-quality model for free, reducing competition.
Offering paid services for companies to fine-tune Llama 3 with their own data.
Selling cloud computing resources for developers to host and run Llama 3 models.

[-]

Infamous-Scallion363@reddit

🧠 AI Knowledge Sharing: I regularly write about the latest AI topics, democratizing relevant AI knowledge.

🌐 Meta's Open-Source Strategy: We'll discuss whether Meta's open-source move is beneficial to the community or strategic like Android's approach to IOS.

📜 Open-Source vs Closed-Source AI: Open-source models are publicly accessible for modification and distribution, fostering collaboration and innovation. Proprietary models are controlled by private companies with exclusive access.

🔍 Meta's Unique Approach: By open-sourcing LLaMA 3, Meta aims to control the platform and drive AI development, positioning itself as an industry leader.

📱 Historical Parallel: Similar to Android's dominance over iOS through open-source, Meta hopes to gain an edge over competitors by open-sourcing its AI models.

🚀 Market Impact: Meta's massive user base on Instagram, WhatsApp, and Facebook could drive quick adoption of Meta AI, integrating it into daily activities.

💰 Monetization Strategies: Plan includes:

Selling access to larger AI models and compute power.
Introducing ads in AI interactions.
Offering generative-AI tools for business messaging.

🏆 Conclusion: Meta’s open-source strategy seeks to drive AI innovation and market leadership with long-term monetization plans. Success will be determined over time.

https://www.linkedin.com/pulse/metas-10-billion-gamble-why-launched-llama-3-siddharth-asthana-dbhze/

[-]

ZaggyChum@reddit

Don't look a gift llama in the mouth.

[-]

wahnsinnwanscene@reddit

To test whether parametric memory is resistant to external efforts of quantization.

[-]

Minute_Attempt3063@reddit

Meta is bad at like the privacy stuff..

But they also love open source.

If they want something better then OpenAi with a worse model, then their best move is to open source it.

Since if more people will use it over OpenAI, they have won already. Sure, bit in the money, or control. But in the name of the people they won.

Meta is one of those companies that are... Very weird with many things.. but i do believe that Zuckerberg is a good person (somewhat) with good ideas. Just the execution is not always good. Or privacy for that matter.. but they are a US company, stealing data is the norm there lol..

This is something more LLM companies need to do. And not let OpenAi create the laws that only benefit them

[-]

galtoramech8699@reddit

How free is it? What are the other top free modells.

[-]

Unhappy-Magician5968@reddit

It forces competitors to keep prices in check against a free option and simultaneously keeps them from a lower quality model l

[-]

karxxm@reddit

Meta has a long history of open sourcing

[-]

MrVodnik@reddit

Very good piece that puts this in perspective of open source, presumably "leaked" from Google.

https://www.semianalysis.com/p/google-we-have-no-moat-and-neither

"Paradoxically, the one clear winner in all of this is Meta"

TL;DR; Meta (and many other companies) did open source many, many projects from their internal operations. They tend to make the exact tool they need, then open source it, and have it maintained for free by the open source community. Even Pytorch that is basically the industry standard was developed by them, which actually took over place of Google's TensorFlow.

This gives them quite a few advantages (other than free labor of course), like having the huge impact on the entire industry tooling development and rapid innovation (many, many independent people doing different changes at once).

[-]