Miau Labs / Insight
Frontier Post-Training Recipe Review with Finbarr Timbers
In this episode of Interconnects, Finbarr Timbers joins Nathan Lambert to discuss the evolution of post-training recipes for language models. They review the key post-training recipes historically, from InstructGPT to the current open frontier models, and explore the emergence of MOPD as a powerful tool for creating do
In this episode of Interconnects, Finbarr Timbers joins Nathan Lambert to discuss the evolution of post-training recipes for language models. They review the key post-training recipes historically, from InstructGPT to the current open frontier models, and explore the emergence of MOPD as a powerful tool for creating domain-specific models. MOPD is a powerful tool for creating domain-specific models, but it's essential to consider the long-term implications and how it may affect the language model industry.
In this episode of Interconnects, Finbarr Timbers joins Nathan Lambert to discuss the evolution of post-training recipes for language models. They review the key post-training recipes historically, from InstructGPT to the current open frontier models, and explore the emergence of MOPD as a powerful tool for creating do
- Finbarr Timbers discusses the evolution of post-training recipes for language models, highlighting the importance of MOPD in enabling the creation of domain-specific models.
- MOPD (Multi-teacher On-Policy Distillation) is a pattern emerging across the 2026 frontier, where multiple domain-specialist teachers are trained and a general student is trained.
- The new frontier models, such as MiMo Flash V2, DeepSeek V4, and Nemotron 3 Ultra, are scaling MOPD to more than 10 teachers, making it organizationally scalable and cost-effective
Miau Labs takeMOPD is a powerful tool for creating domain-specific models, but it's essential to consider the long-term implications and how it may affect the language model industry.