Meta has a moat. That moat is the connected graph of users across its various properties, and the (monetisable) attention of those users. Making LLMs a commodity reduces the likelihood that OpenAI (or someone else) will build a unique, unreproduceable product that rivals Meta's properties for attention.
Unless the existence of Llama somehow inhibits users from downloading and using other LLMs (a-la Microsoft preinstalling IE) then I don't see how any just society would implicate them. Even if Llama becomes a monopoly; what's the market abuse in making something free?
Using a monopolistic position in one market (i.e. cash flow) to make another market not profitable can be considered abusing one's position in the first market.
i.e. can be compared to a form of "dumping" (not quite what dumping is normally, but close enough I think0.
1) Talent. The more engineers are familiar with Llama, PyTorch, and the likes, the easier it is to find ML talent for Meta.
2) Free research and innovation that Meta can readily copy for their needs.
3) They don’t really give up any power. They can always keep future models proprietary.
In my opinion, it is a bit like Microsoft’s “Embrace, Extend, Extinguish.” Make SW publicly available, extend it to your own business needs, and if a competitor emerges, the power is still with Meta to restrict its use or copy the use case (“extinguish”).
Having thousands of people working on making your models better is pretty nice. And it undercuts their competitors who rely on income from LLM services rather than selling people like Facebook does.
I get what you mean, but let's just clarify what happens with free software: thousands of people are eligible to contribute hassle-free if they want to, some people help by filing bugs and minor fixes, and maybe sometimes a few people actually contribute great things you wouldn't have managed to do and the software may thus become much better.
PR is the research community. More top researchers want to work where their outputs have a wide reach and impact.
From the outside, it seems LeCun is most interested in the science and progress of ML/AI and has convinced Zuckerberg that open is also good for business. I’m not sure if that’s true, but it might be.
The reasoning I’ve heard voiced by Zuckerberg is that the faster the field moves forward, the better they can make Facebook and their other products. Their product isn’t models, it’s social media. Better data science helps them make better social media
> What does Facebook gain in making LLAMA freely available?
The business model that consists in selling a subscription to be able to query a LLM isn't much of one: it's a race to the bottom. Users will switch as soon as anything even slightly better is out (like cancelling ChatGPT and taking a Claude 3 Opus subscription), companies with similarly performing models have to compete on price. It's ugly.
By making models freely available and by buying and advertizing that they bought a shitload of NVidia GPUs to train AI models (LLMs and others), Meta is adding additional pressure to those already fighting in an overcrowded market.
They don't do it because they'd suddenly be nice: Meta is still a despicable company with despicable tactics. It's all about screwing these one-trick ponies as much as they can: last thing Meta would want would be a tiny player to become a trillion behemoth.
It's an unfair competition though. Funding sources, data sets, hardware availability; it's like holding a race to get from point A to point B with no rules, so some people are running on foot, some are in cars, some have laid down tracks so they can run a train, others are using a rocket engine that someone else bought them.
Meta has a moat. That moat is the connected graph of users across its various properties, and the (monetisable) attention of those users. Making LLMs a commodity reduces the likelihood that OpenAI (or someone else) will build a unique, unreproduceable product that rivals Meta's properties for attention.