Mark Zuckerberg’s Open Source AI Play

Markdown Version of the Article

Introduction

Some people see you as an unlikely champion of Open Source today. You're laughing.

Well, I don't know why I—I mean, I actually think—well, I get it, but...

You understand the word "unlikely".

Yeah. We've actually, at Meta, been pretty big proponents of open source for a while, and it's been a really good formula for us. Going back to how we designed servers and our data centers, we have this whole Open Compute Project that standardized a lot of the infrastructure for the industry. That's going to happen with AI too.

You're really putting a stake in the ground by open-sourcing your AI in this attempt to build the AI rails for the future. How much of this is a strategic way to control or own the next technology wave?

A lot of our growth over the last 10 or 15 years was building our apps through phone platforms controlled by competitors. It's somewhat soul-crushing to build something you think is going to be good and then be told by Apple that you can't ship it because they want to put us in a box as they view us as competitive. We're a big enough company now that I've resolved for the next generation of technology; we want to build and have more control over the next set of platforms that we're going to build. AI is a critical one, and augmented and virtual reality is another. The way we can control our own destiny and ensure we have access to leading AI is by building it and having it become an industry standard. It actually gets stronger by sharing it and having the ecosystem continue to improve it.

You're continuing to improve Meta AI across all of your products but also as a standalone chatbot. Why should we use Meta over ChatGPT?

There are a bunch of things where Meta AI is better. One of the things we're rolling out soon is the ability to just imagine stuff as you're typing in real-time. I do this with my daughters all the time. As you're entering the query, it's generating images as you enter the keystrokes. It's really cool. My goal for the Meta AI launch, which is only a few months old at this point, was to make Meta AI the most used AI assistant in the world by the end of the year. I think we're on track for that, with hundreds of millions of people using it by then.

You’re releasing LLaMA 3.1, this family of models, big and small, including the biggest open-source model ever: 45 billion parameters. What does that jump unlock?

I think the main thing people will do, especially because it's open source, is use it as a teacher to train smaller models for different applications. Think about all the startups, enterprises, or even governments trying to do different things. They all probably need to build custom models for their work, and it's really hard with closed systems out there, whether that's OpenAI, Gemini, Google, or others. This aligns with our core philosophy: we don't believe there will be one AI to rule them all. Our vision is for millions or billions of different models out there.

So not one God but many—is that the way to think about it?

I don't think there will be gods, but if you're an organization thinking you will create one super-intelligent AI, it does have a feel of trying to create a god. I find that approach both wrong and unappealing. We have almost 200 million creators on our platforms, all trying to build their communities. People want to interact with them, but there aren't enough hours in the day. I want to make it so that every creator can easily train an AI version of themselves, like an artistic artifact that they're putting out for their community, allowing interaction but also controlling how that interaction happens.

Facebook and you have been blamed for a lot of things, whether you agree or not. Why should we trust you with AI?

It's a loaded question. We have gotten blamed for a lot of things, and I take our role in all this seriously. We've tried to handle all this as well as possible. I'm not sure it's all been fair, but I like to think we're an important and relevant company, so the scrutiny is generally healthy. One of the defining things around open source is that anyone can scrutinize the work, which puts a lot of pressure to ensure the quality of what you're doing gets better quickly.

Keywords

• Mark Zuckerberg
• Open Source AI
• Meta
• Open Compute Project
• AI Industry Standards
• Meta AI vs. ChatGPT
• LLaMA 3.1
• Custom Models
• AI Ecosystem
• Trust

FAQ

Q: Why does Mark Zuckerberg think Meta should open-source its AI? A: Zuckerberg believes that open-sourcing AI strengthens the technology by allowing the ecosystem around it to continue improving it, and it is also aligned with their philosophy of having numerous AI models rather than a single dominant one.

Q: What distinguishes Meta AI from competitors like ChatGPT? A: Meta AI offers features such as real-time image generation as you type, and Zuckerberg aims to make it the most used AI assistant in the world by the end of the year.

Q: What is the significance of LLaMA 3.1 and its 45 billion parameters? A: It is the largest open-source AI model ever released. It serves as a teacher for startups, enterprises, and governments to train smaller, custom models for various applications.

Q: How does Zuckerberg envision the future of AI models? A: He envisions millions or billions of different AI models tailored for various applications, opposing the idea of a single, super-intelligent AI.

Q: Why should people trust Meta with AI, given their controversial past? A: While acknowledging past criticisms, Zuckerberg emphasizes the benefits of open-source development, where the transparency allows for widespread scrutiny and continuous improvement of the technology.