Meta CTO Andrew Bosworth admitted the company's Llama 4 development broke its own training pipeline after overextending resources on Llama 3, and is now renting frontier models from competitors including OpenAI, Anthropic, Apple, and Google to fill the gap. On the same day as the interview, Meta released Muse Spark 1.1, a new agentic model it claims is benchmark-competitive at lower cost than rivals.
Bosworth's core argument is that the monolithic frontier model era is ending and that product, distribution, and consumer experience now matter as much as raw model capability. He names all four elements as Meta's combined advantage, and claims no single competitor holds more than one of them. That framing deserves scrutiny, and the full interview presses it.
The conversation covers Meta's wearable hardware roadmap, its target of 'personal superintelligence,' and the internal cultural strain of keeping pace with the current AI cycle. The tension between Bosworth's confidence in Meta's position and the reality of renting models from the companies he's competing against is the thread worth pulling.
[READ ORIGINAL →]