Sebastian Raschka's from-scratch video course on building reasoning models
magazine.sebastianraschka.com ·
An eight-video, 11-hour coding course teaches developers to build and improve a small reasoning model in Python and PyTorch, beginning with a pre-trained Qwen3 0.6B model. The materials cover model loading, streaming text generation, KV caching, model compilation, evaluation, inference-time scaling, reinforcement learning, and distillation. Python and PyTorch familiarity are recommended; early generation and evaluation exercises can run on a CPU, while later training experiments are more practical on a GPU.
The course builds a math-answer verifier and uses the MATH-500 benchmark to evaluate models. It implements temperature scaling, multinomial sampling, top-p filtering, self-consistency voting, scoring based on token and sequence log probabilities, and self-refinement. Later lessons implement a simplified Group Relative Policy Optimization method for reinforcement learning with verifiable rewards, including response groups scored by the verifier, training metrics, clipped policy ratios, and KL-loss terms. The final lesson distills reasoning behavior from DeepSeek R1-generated answers through supervised training and compares distilled-model results on MATH-500. The videos accompany the book Build a Reasoning Model (From Scratch), and a separate 90-minute LinkedIn Learning course provides a conceptual overview of inference-time scaling, reinforcement learning, and distillation.
What people are saying
The thread spent far more time on whether a from-scratch LLM course is worth taking than on reasoning-model techniques themselves.
What readers liked
- The instructor’s step-by-step treatment makes backpropagation, autograd, and LLM training feel like connected machinery rather than opaque…
- Readers praise the books and videos for turning dense neural-network material into incremental exercises where each stage produces a tangible advance.
Concerns raised
- Skeptics doubt that a video course can prepare anyone for the real cost, data demands, and technical depth of training an independent model.
- The course risks serving people who want the identity of building AI more than a realistic path to deploying one, especially if hands-on work never…
Read full discussion analysis
Source: Hacker News · Based on 10 sampled comments · Analyzed
Discussion overview
The thread spent far more time on whether a from-scratch LLM course is worth taking than on reasoning-model techniques themselves. Supporters argued that the premise is valuable precisely because it is not a shortcut to founding a frontier lab: rebuilding the pieces teaches attention, backpropagation, PyTorch, and training mechanics in a way that black-box API use never will. Readers repeatedly pointed to the instructor’s books as proof that the material can turn neural-network concepts from simultaneously mystical and tedious into a sequence of understandable steps. Study groups and hands-on work were also seen as making a difficult subject less isolating.
The main skepticism was practical. One reader questioned whether video lessons serve anyone beyond hobbyist understanding: people serious about training their own models need much deeper preparation and substantial resources, while those who only want conceptual literacy may not need an extended build course. Another reader described the familiar ML-learning trap: theory can demand near-mathematics-degree commitment, but meaningful applied experiments quickly run into compute and data barriers. Optimists pushed back that small teams may still build narrow, specialized models even if they cannot match frontier systems on broad benchmarks. The dominant mood is enthusiastic about the teaching, with a real but minority concern that accessibility may stop at toy-scale exercises.
Points of disagreement
The sharp disagreement is whether a build-from-scratch reasoning-model course is substantive training or mostly aspirational education. Skeptics argue that real model-building requires resources and depth a video series cannot provide, making it easy to sell ambition without practical outcomes. Supporters argue that understanding the machinery is itself a legitimate goal, and that specialized small-team models could make this knowledge practically useful.
What readers liked
- The instructor’s step-by-step treatment makes backpropagation, autograd, and LLM training feel like connected machinery rather than opaque mathematical ritual.
- Readers praise the books and videos for turning dense neural-network material into incremental exercises where each stage produces a tangible advance.
- Learning by rebuilding an LLM is defended as valuable even for non-founders, because it gives people a working mental model of technology otherwise encountered only through APIs.
- Several commenters see room for small teams to use these foundations on narrow domain models, where specialization may matter more than winning broad frontier benchmarks.
Concerns raised
- Skeptics doubt that a video course can prepare anyone for the real cost, data demands, and technical depth of training an independent model.
- The course risks serving people who want the identity of building AI more than a realistic path to deploying one, especially if hands-on work never escapes toy exercises.
- ML education remains caught between theory that feels academically forbidding and practical projects that become inaccessible once serious compute is required.
Keep reading
Sources and details
magazine.sebastianraschka.com · Published
Sebastian Raschka's from-scratch video course on building reasoning models
Read report →
Community analysis available for 1 of 1 linked discussions.
Additional discussion searches are still incomplete.
Coverage details and dates
Latest available community analysis:
Summaries: 1 current · 0 awaiting update · 0 not verified
Latest summary update:
Hacker News · magazine.sebastianraschka.com