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

Original Article ↗ Hacker News Discussion ↗

Sebastian Raschka’s new free video course on building “reasoning models” – LLMs enhanced with reinforcement learning and verifiable rewards – is widely praised for making complex AI concepts approachable, especially as a follow‑up to his popular “LLMs from scratch” book. Commenters debate who this kind of material is really for: some see it as essential groundwork for engineers, researchers, and serious hobbyists who want to understand and fine‑tune models, while others question its practical value given the compute and data barriers to training large systems. Many argue that small‑scale experiments, post‑training open models, and domain‑specific applications still offer substantial room for meaningful hands‑on learning, even without building frontier‑scale AI.

Overall reception of the course and books

  • Many commenters strongly praise the books and new video course as clear, intuitive introductions to LLMs and reasoning models.
  • Several people say concepts like backpropagation, attention, and PyTorch finally “clicked” after using these materials.
  • Multiple readers call the LLM book an ideal first step, with the reasoning-model material as a natural second step.
  • Some note the resources are substantial enough to be useful even to people already training models professionally.

Target audience and purpose

  • One thread questions who this is “for,” suggesting it mainly sells an aspirational identity to people who won’t actually build models.
  • Others push back, arguing:
    • The videos are free; monetization fears are overstated.
    • Learning-by-doing at small scale is valuable even if you’ll never train frontier models.
    • This is analogous to learning guitar or baking: you don’t need to become world-class for the learning to be worthwhile.

Hobbyist feasibility and compute requirements

  • Several commenters feel ML/LLMs are less “practical” to learn than general software because of heavy data/compute needs and messy real-world data.
  • Others argue:
    • You don’t need huge models; small models and classical ML remain useful.
    • Fine-tuning and post-training existing open models (e.g., via LoRA, RL, or domain-specific tasks) is accessible and meaningful.
    • The book/course code is designed to run on a regular laptop; a consumer GPU only speeds things up.

Practical next steps after learning

  • Suggestions include:
    • Fine-tuning small models on personal or niche datasets.
    • Building small classifiers for personal finance, email, or photo organization.
    • Contributing to open-source tooling (inference engines, training frameworks).
    • Exploring ML in other domains like bioinformatics or genomics.

Debates on “build your own model” guides

  • Some compare these guides to “build your own CPU” tutorials: educational but not practically useful.
  • Others counter that such knowledge:
    • Is excellent interview preparation.
    • Provides essential intuition about training, optimization, and RL-based post-training.
    • Is a foundation for specialized or hyper-focused models where small teams can still innovate.

Conceptual questions: reasoning models & RLVR

  • A late thread questions the definition of “reasoning models,” their relation to multi-path inference, agents, and reinforcement learning with verifiable rewards.
  • Timelines and mechanisms for how these pieces fit together are actively debated and remain somewhat unclear in the discussion.