How Zuckerberg rebuilt Meta around Llama

Meta's Llama models have now logged over 600 million downloads on open-source hubs like Hugging Face, turning into the default foundation for the open AI community.

More importantly, Llama has become the technical backbone of Meta itself, powering everything from conversational assistants on Facebook, Instagram, WhatsApp, and Messenger to Ray-Ban smart glasses and ad-creation tools.

A report published on November 19 pulled back the curtain on how Mark Zuckerberg broke an internal executive deadlock to push Llama 2 out the door and reshape the company.

A messy road to open source

Back in the summer of 2023, Meta's leadership was split on whether to open-source Llama 2.

Chief AI scientist Yann LeCun and FAIR head Joelle Pineau lobbied hard for a full open-source release, arguing it was the fastest and cheapest way to catch up with OpenAI after Meta stumbled in the generative AI race.

Other executives pushed back hard. Giving away cutting-edge software made direct monetization nearly impossible, and releasing model weights raised terrifying legal and cybersecurity liabilities.

Meta had already been burned. In late 2022, its Galactica scientific model was pulled after just three days because it generated wild falsehoods, including instructions on making napalm in bathtubs and eating crushed glass.

When the original LLaMA leaked online shortly after, Meta's distribution security was compared to Swiss cheese by global affairs chief Nick Clegg, drawing a sharp letter of reprimand from two U.S. senators.

Ahmad Al-Dahle, Meta's head of generative AI, recalled that Zuckerberg talked to stakeholders on every side of the issue before finally making the call himself on July 18, 2023: "Okay, we're just going to do it."

Powering the product line

That single gamble helped Zuckerberg shake off the metaverse hangover, where legless avatars and endless spending had hammered Meta's stock to seven-year lows.

By leveraging billions of publicly shared posts across Instagram and Facebook for training, Meta quickly closed the gap. When Llama 3 rolled out across April and July 2024, its largest model matched or beat proprietary benchmarks from OpenAI and Anthropic.

Today, thousands of enterprise giants like Goldman Sachs, AT&T, and Accenture build on Llama for free.

Inside Meta, Llama drives AI Studio chatbots, Quest headset vision tools, ad generation engines, and the popular Ray-Ban Meta glasses.

Yet the massive price tag keeps investors nervous. On its third-quarter earnings call, Meta projected 2024 capital expenditures reaching up to $40 billion, with another significant leap coming in 2025. Abhishek Nagaraj, an associate professor at UC Berkeley's Haas School of Business, noted that the move remains hard to justify from a purely economic standpoint.

Why it matters

Zuckerberg often compares Llama to Linux, and we think the comparison reveals his real motivation.

Meta spent the last decade suffering under the mobile app store rules imposed by Apple and Google. By making frontier-tier AI free and open, Meta is trying to ensure it never gets locked out of the next computing platform, while simultaneously commoditizing the software moats of rivals like OpenAI and Google.

For developers and startups, Llama is an incredible gift that eliminates vendor lock-in and slashes the cost of running intelligent applications.

Still, spending $40 billion a year on infrastructure while giving away the core engine is an enormous gamble. Unless ad targeting and smart hardware sales deliver massive windfalls, Meta may eventually face tough questions from Wall Street over how long it can fund open source out of its own pocket.