Reasoning models taught from scratch

Sebastian Raschka has published eight coding videos, more than 11 hours in total, on building a reasoning model from the ground up. The course starts with a pre-trained Qwen3 model and shows how to evaluate and improve it using Python and PyTorch, covering inference-time scaling, reinforcement learning and distillation. It drew 353 points and 52 comments on Hacker News, a sign of strong interest in learning how these systems work, not proof of how widely companies are adopting them. Raschka’s course is aimed at people comfortable with Python and PyTorch.

For an operator, one useful takeaway is the distinction between a reasoning model and an agent. Raschka describes the model as the engine, while an agent harness provides tools and manages context so the system can handle a task. That’s a practical way to think about AI products: a model’s reasoning ability is only one part of what makes a tool useful for work.