"Optimize for inference cost when users massively outnumber trainers Meta made the unusual decision to train Llama 3 on far more data than is theoretically optimal for a single training run, because their serving scale inverts normal priorities: "our ratio of inference compute required to training is probably much higher than most other companies that are doing this stuff just because the sheer volume of the community that we're serving." Even with the 70-billion model, "by the end it was still learning" — they stopped not because the model stopped improving but because they needed to move on. Whenever you deploy something to millions of users, optimize for the cost of running it at scale, not just the cost of building it. Source: Mark Zuckerberg, Dwarkesh Patel Podcast, Llama 3, Meta AI, Future of AI"