How Claude works: LLMs, alignment, interpretability, agents

11 steps · about 7 h 30 min · Intermediate

  1. Large Language Models explained briefly (video, 3Blue1Brown, 8 min)

    A short, visual picture of next-word prediction, pretraining and chatbots gives you the base for everything after it.

  2. Transformers, the tech behind LLMs | Deep Learning Chapter 5 (video, 3Blue1Brown, 27 min)

    This shows how data flows through a transformer, which you need before reading about training or interpretability.

  3. Anthropic's Prompt Engineering Interactive Tutorial (interactive, Anthropic (GitHub), 2 h)

    Hands-on exercises show you how a model responds to instructions, so you learn its behaviour by using it.

  4. Prompt engineering overview (doc, Claude Platform Docs, 15 min)

    It adds the current official guidance, including defining success criteria and testing prompts.

  5. Claude's constitution (article, Anthropic, 15 min)

    A plain-language overview of Constitutional AI, the idea of training a model on written principles.

  6. Constitutional AI: Harmlessness from AI Feedback (article, arXiv (Bai et al., Anthropic), 1 h)

    The original paper covers self-critique, revision and reinforcement learning from AI feedback (RLAIF). Skim the method sections after the overview.

  7. Building effective agents (article, Anthropic, 25 min)

    This turns the model into systems: workflows versus agents, and why simple, composable patterns tend to win.

  8. Writing effective tools for agents — with agents (article, Anthropic Engineering, 25 min)

    It shows how tool design and evaluation decide whether an agent works.

  9. Tracing the thoughts of a large language model (article, Anthropic, 20 min)

    A readable tour of what interpretability research has found inside a model, such as planning ahead and shared concepts across languages.

  10. Open-sourcing circuit-tracing tools (interactive, Anthropic, 45 min)

    You can generate and explore attribution graphs on open-weights models yourself instead of only reading about them.

  11. Transformer Circuits Thread (reference, Anthropic Interpretability team, 1 h 30 min)

    The primary research papers, including Towards Monosemanticity and Scaling Monosemanticity, are the final step for going deeper.