Podcast, The Information Bottleneck
Post-Training Is Where Models Are Made Now | Maxime Labonne
Hosted by Ravid Shwartz Ziv and Allen Roush
About
Maxime Labonne, head of post-training at Liquid AI and author of the LLM Course, traces post-training from supervised fine-tuning and preference optimisation to reinforcement learning with verifiable rewards and multi-domain on-policy distillation, in which separate code, maths and other domain teachers are merged back into one student model. He explains why supervised fine-tuning cannot be skipped, why stage order matters, why reinforcement learning works although a long rollout yields about one bit of signal, and his view that fine-tuning and distillation raise pass@k while reinforcement learning sharpens it into pass@1. The conversation also covers training small agentic models inside their harness, open-source reinforcement-learning stacks, looping transformers, why models now write worse, safety and regulation, open and Chinese models, and whether Europe can build frontier models.
Topics
- post-training
- supervised fine-tuning
- preference optimisation, broader topic: optimisation
- Reinforcement Learning
- verifiable rewards
- on-policy distillation, broader topic: distillation
- pass@k, search World Wide
- pass@1, search World Wide
- agentic models, search World Wide
- looping transformers, broader topic: transformers