“Gains from pre-training AI models began to asymptote approximately 18 to 24 months ago as labs had largely ingested all available knowledge from the internet.”