Meta fell behind in the frontier AI race. Andrew Bosworth, Meta's CTO, admits it directly and explains how Llama lost ground, why it happened, and what Meta is doing to recover. This is not a PR interview.
The argument Bosworth makes is structural: the best AI products will not run on a single model. Multi-model architectures are the direction, and Meta's distribution across glasses, Ray-Ban hardware, and consumer apps is the lever it plans to use against OpenAI and Google. He also addresses renting versus owning frontier capabilities, which has real implications for how Meta prices and positions its AI stack. Consumer agents have not broken through yet, and Bosworth gives a specific account of why.
The parts worth reading in full are his candid breakdown of what went wrong with Llama, his framing of employee tracking and internal training programs as necessary friction in a technological transition, and his position on AI companions as a serious product category. The conversation is uncomfortable in the right places.
[WATCH ON YOUTUBE →]