About the method

This course decomposes twelve source infographics on LLM systems into 120 teaching plates across fifteen lessons. Each plate presents one concept through eight anchors: orientation, mental model, mechanism, interactive exploration, implementation, failure modes, lab, and takeaways.

Source decomposition

The original infographics communicated dense ideas in a single frame. This course takes each one apart: identifying what is correct, what needs qualification, and what must be corrected. Every correction is documented and traceable to a specific claim in the source material.

The decomposition follows a consistent structure: one causal idea per plate, progressive disclosure of detail, and a strict boundary between what the model does and what the application does. No plate introduces a concept that depends on a later plate.

The MiraForge voice

Content is written in the voice of a master craftsman explaining their workbench to a capable apprentice. Technical accuracy is non-negotiable. Analogies illuminate where they can and state where they break. The forge metaphor appears throughout the course, particularly in the module framing and the mental models for sampling, retrieval, and agent loops.

Visual identity

All artwork is generated with GPT Image 2 using the Mirai Forge Technical Atelier visual language: warm fibrous ivory paper, charcoal sumi ink contours, muted indigo structure, one restrained vermilion focal point, tiny brushed-brass details, soft raking morning light. Japanese influence comes from materials, space, and craft — not costume or stereotype.

Native educational diagrams (equations, data tables, code blocks) are built as live HTML, not baked into images. Generated artwork provides atmosphere and orientation; precise labels and data remain editable, accessible, and responsive.

Editorial standard

Every paragraph states a falsifiable claim, explains a mechanism, marks a boundary, interprets evidence, specifies an action, or names a failure. The course names the actor at every step: the model, the host application, external data, a tool, an evaluator, or a human. It does not write that AI remembers, understands, decides, or knows when the actual actor is something specific.

Module 0: Foundations

The original course opened directly with token mechanics, assuming ML vocabulary the audience does not have. Module 0 was added to build the foundations: what a neural network is, why the transformer architecture matters, and how training turns architecture into what we call an LLM. The curriculum was validated by an ML subject matter expert and revised to address three critical gaps in the original design.