“Amjad Masad observes a shift in LLM development from pure pre-training scaling to a "functional AGI" approach, which involves creating specialized reinforcement learning environments and collecting domain-specific data to automate tasks in fields like coding and investment banking.”