Sebastian Raschka's from-scratch video course on building reasoning models

771 points · 104 comments on HN · read original →

Points and comments are a snapshot, not live.

Sebastian Raschka released 8 free video lectures (11 hours) on building reasoning models from scratch.

Raschka's course uses a pre-trained Qwen3 0.6B model and implements inference-time scaling, reinforcement learning (RLVR with GRPO), and distillation in PyTorch. It builds a math verifier for MATH-500, implements self-consistency, scoring, and self-refinement. The 8 videos accompany his book 'Build a Reasoning Model (From Scratch)'. He distinguishes reasoning models (the engine) from agent harnesses (the tooling framework around the model).

What commenters are saying

Strongly positive reception: commenters praise his pedagogy, calling the book 'super readable' and 'almost perfect.' Several note that while some home-hackers find ML impractical due to compute requirements, post-training existing small models via LoRAs is a practical path. One commenter questions the target audience, suggesting it's a personal-brand monetization strategy, but others defend learning-for-learning's sake. No substantive corrections to the article appear.