How Sebastian Raschka Teaches Reasoning Models From Scratch

What does the course teach?

Sebastian Raschka’s course, “Build Your Own Reasoning Model,” is a set of eight coding videos lasting more than 11 hours. It begins with a pretrained large language model, then demonstrates ways to evaluate and improve it using Python and PyTorch. ¹

The subjects include evaluation, inference-time scaling, reinforcement learning and distillation. Inference-time scaling means using additional computation while a model is responding; distillation is a way of transferring capabilities from one model to another. The course description identifies these topics, but does not provide enough detail to assess every lesson’s depth or exercises. ¹

How can you build a reasoning model from scratch?

Here, “from scratch” describes implementing methods in code to understand how they work. The course does not say that learners train a large language model from entirely untrained parameters. Instead, it starts with a model that has already been pretrained and shows how to evaluate and improve it. ¹

That distinction matters. Pretraining a model and building methods around a pretrained model are different tasks. Raschka’s description is about the latter: coding and testing approaches that may help a model reason more effectively. ¹

What concepts should learners understand?

Evaluation is the process of checking how a model performs on tasks. It helps learners compare versions or approaches rather than relying on an impression that a model seems better. The course lists evaluation as one of its main subjects. ¹

Reinforcement learning is a way to train systems using feedback about their actions. Distillation transfers useful behaviour or capability from one model to another. Inference-time scaling explores how giving a model more computation while it generates an answer affects its output. These are broad descriptions; the available course summary does not specify all the techniques or training setups covered. ¹

More generally, reasoning in AI can involve representing information in a form a computer can use to work through tasks. That broad idea is not the same as a detailed account of how a particular language model’s reasoning works. ⁴

Do you need to take Raschka’s earlier course first?

According to Raschka, the reasoning-model videos can be watched without first completing his earlier “Coding LLMs from the Ground Up” course. He presents the new material as a sequel, but says it does not require that earlier course as a prerequisite. ¹

The lessons use Python and PyTorch, so learners should expect to follow code rather than only watch general explanations. The source does not give a formal list of required programming skills, provide a syllabus for each video, or say what hardware is needed. ¹

What can learners take away, and what remains unclear?

The course offers a way to study several methods by implementing them. That can make technical ideas more concrete than a description alone, but the source does not report independent assessments of the course or evidence that completing it guarantees a particular skill level or model result. ¹

A separate educational project offers an example of a different approach to reasoning models: an implementation of a hierarchical model applied to finding paths through a board with obstacles. It describes two recurrent modules operating at different frequencies and includes a performance study, but it is not part of Raschka’s course. ²

The practical takeaway is to treat the videos as a coding course on methods for evaluating and improving a pretrained model, not as a promise that learners will produce a state-of-the-art system. The available description leaves details about lesson-by-lesson content, costs, access conditions and outcomes unanswered. ¹