How frontier AI models are moving toward multi-teacher training recipes

The piece examines the evolution of post-training recipes—techniques used to refine language models after their initial training phase. Through a discussion between Nathan Lambert and Finbarr Timbers, both specialists in AI model development, the content traces the progression from InstructGPT (2022) to the frontier models of 2026. The dominant pattern emerging in 2026 is Multi-teacher On-Policy Distillation (MOPD), a method where multiple domain-specialist models teach a single general model. This approach arose because traditional reinforcement learning became computationally expensive and prone to capability conflicts when optimizing for mathematics, code generation, and agentic behavior simultaneously. The analysis includes a historical review spanning from InstructGPT's simple pipeline (SFT → reward model → PPO) through Llama 3's multi-stage approaches, to recent models such as DeepSeek V4, MiMo Flash v2, and Nemotron 3 Ultra, which have adopted the MOPD strategy. The discussion also addresses organizational challenges: advanced post-training is not merely a technical problem, but one of organizational capacity to coordinate compute resources and talent across research teams.