I agree that LLMs are not a complement - Facebook is not an organisation that desparately needs a bunch of LLM-generated content.
They have masses of content generated for free by users and journalists and influencers and so on - if anything, a bunch of LLM spam is a threat to that.
However, Open-weights LLMs are a much smaller threat to Facebook than they are to Google (where it could replace a lot of search usage) or Open AI (whose business is selling LLM access)
Perhaps for Facebook the benefits of the open weights approach - where you give away the model and get back a load of somewhat improved models, a faster way of running it, and a load of experienced potential hires - pays off because it doesn't threaten their core business.
> Facebook is not an organisation that desparately needs a bunch of LLM-generated content.
This is an overly narrow view of what an LLM can do. Generating text is the really neat parlor trick that people are trying to cram in to every possible startup, but if you take a broader view then what LLMs really are is the single largest breakthrough in natural language understanding.
Facebook doesn't need text generators, but they do need language understanding, especially for recommendation and moderation.
I'm not convinced that it's a complement—Joel's explanation is that you make a product that users consume alongside yours very cheap in order to keep people coming to you— but they definitely need LLMs.
Meta's business model is figuring users out and selling ads to them, as well as having to police posts on an industrial scale to try to remove stuff like election interference, terrorist videos, etc. AI is used for all of this.
The GPU cluster that they trained their Llama models on was actually built to train Reelz (their TikTok competitor) to recognize video content for recommendation purposes, which is the thing that TikTok does so well - figuring out users' preferences.
While this is true, that doesn't make them a complement to Facebook's broader business. Here's Joel's definition:
> A complement is a product that you usually buy together with another product. Gas and cars are complements. Computer hardware is a classic complement of computer operating systems. And babysitters are a complement of dinner at fine restaurants.
LLMs aren't really a complement like gas to cars because the end user doesn't need to consume the LLM in order to use the social media site. It's more like LLMs are becoming an essential component of a social media site—not like gas to cars but like an engine control unit, a part that ideally the user will never see or interact with. Joel's reasoning doesn't apply to that kind of product because users don't see the price of LLMs as a barrier to consumption of social media.
That doesn’t make it a compliment. A compliment is what a facebook advertiser or facebook user would also buy (or at least buy with their time) along with facebook. 5G data might be an example for FB users.
There are many ways. Generative AI helps people create content (and that's not the only way, I'm sure). Meta's platforms use content to drive attention.
For instance, an Instagram account that shows cool AI generated photos generates ad revenue for Meta.
They have masses of content generated for free by users and journalists and influencers and so on - if anything, a bunch of LLM spam is a threat to that.
However, Open-weights LLMs are a much smaller threat to Facebook than they are to Google (where it could replace a lot of search usage) or Open AI (whose business is selling LLM access)
Perhaps for Facebook the benefits of the open weights approach - where you give away the model and get back a load of somewhat improved models, a faster way of running it, and a load of experienced potential hires - pays off because it doesn't threaten their core business.