Interview, Fireside Chat
Mark Zuckerberg — AI will write most Meta code in 18 months
- The global landscape is expected to become increasingly "funny," "weird," and "quirky" as individual cultural content production rises, with digital consumption shifting from passive video to interactive dialogue and a future "intelligence explosion" paced by hardware buildout rather than algorithmic breakthroughs.
- Demand for human labor is projected to increase rather than decrease in sectors where AI reduces service provision costs to one-tenth, with historical automation trends suggesting that cost reductions may ultimately require more staff, such as in customer support, where voice support could become economically viable within a three to five year horizon.
- Meta anticipates that "most" code for Llama research efforts will be written by AI agents within the next 12 to 18 months, while internal "Metamate" agents are expected to accelerate Llama research and software engineering.
- Meta plans to release the "Little Llama" (2B parameter) model in the coming months to drive significant AI usage growth, with future "Llama 4" reasoning models expected to lead multimodal capabilities, though it remains uncertain if the company will continue building its own large models if the open source ecosystem becomes sufficiently healthy.
- Open source models are forecast to generally overtake closed source models in usage, while the industry moves toward specialization in domains like enterprise, coding, productivity, social, entertainment, and companions.
- User interaction will evolve toward co-evolution over a year or two, where assistants learn preferences as people learn to use them, eventually enabling all-day communication via phones, feed apps, and messaging apps, with future devices like glasses designed to "get out of the way" rather than cluttering views.
- Meta is developing a "full duplex voice" feature for its new AI app, though it is not yet the default, and the company plans to offer arbitrary compute amounts for "really even more amazing things" through a premium service model that coexists with free, ad-supported tiers.
- The future business model will likely balance free ad-supported services with premium tiers for high-compute tasks involving software engineering agents, while AI development shifts to being "model led" rather than "product led."
- Societal trends predict a long-term decline in energy spent on physical needs over a 100 to 150 year arc, increasing time for creative pursuits, and a cultural shift where interacting with AI companions or therapists is viewed as rational and valuable.
- Significant infrastructure risks include physical bottlenecks in building gigawatt clusters and potential U.S. disadvantages compared to China if energy power and permitting are not streamlined, alongside a transition to a world where AI becomes the next generation of search for complex information tasks.
- Recruitment remains the primary leveraged activity for leadership, followed by resolving cross-team product integration, as the trajectory emphasizes that reducing costs often leads to expanded human demand rather than replacement.
- Meta's current strategy assumes the need for custom control over model architecture to compete with others, with internal coding agents advancing Llama research significantly in the next 12 to 18 months.