Mark Zuckerberg — AI will write most Meta code in 18 months

Zuckerberg describes Llama 4 as a staged release rather than a single model launch. Meta announced four models and initially released Scout and Maverick, aiming for efficient, low-latency, natively multimodal systems that can run on one host. The company is also preparing an 8-billion-parameter model, internally called Little Llama, and Behemoth, a frontier-scale model with more than 2 trillion parameters. Behemoth is large enough that Meta had to build infrastructure to post-train it, leaving open how developers will use it or whether its capabilities can be distilled into practical models.

The interview begins with a broader product ambition: Meta AI already has nearly a billion monthly users, and Zuckerberg expects personalization to become more important as systems combine a person's feed, profile, social graph, and direct interactions with AI. That vision depends on continuing to build the model family as well as learning from user behavior; it is a company strategy and forecast, not an independently measured outcome in the conversation.

Dwarkesh raises the impression that open models have fallen behind closed models, pointing to Maverick's Chatbot Arena position and benchmark results. Zuckerberg counters that open source has become more diverse and predicts it may become the most widely used model category. He also sees increasing specialization: reasoning models spend more inference-time compute on difficult tasks, while many consumer uses depend more on speed and intelligence per dollar. Meta is developing a Llama 4 reasoning model but expects both approaches to matter.

Zuckerberg is skeptical that public benchmarks fully measure user value. He says tests can focus on tasks ordinary users rarely perform and can be gamed: Meta could tune a version of Maverick to rank highly, while the released model was not tuned for Arena. Meta therefore says it will optimize for reported user value and behavior in its own product, and may compare other models directly. That approach ties evaluation to the company's applications, but also makes its definition of success less independent and less directly comparable across labs.