Guess Wrong First, Then Get Hands-On: This Open-Source Skill Turns Transformer into Actionable Lessons

AI-generated courseware is nothing new. What is hard is generating a lesson that lets people verify things by hand. Teachable Figures recently went open source; it is positioned as a skill for AI agents: turning abstract concepts into interactive 2D and 3D teaching figures. The project is at GitHub, and the online demo is here.

In the post, the author says the starting point was very specific. While learning AI model architecture and AI infrastructure, and reading about systems like SGLang, the author always wanted to see what was happening: where data goes, how components work together, and what changes if you modify one thing. The first version of the skill was a visualization explanation for understanding SGLang. Later, the author saw Anatomy and iso-figure's rendering methods for mechanical structures, cross-sections, exploded views, and isometric functional objects, combined these methods with their own teaching process, and that led to Teachable Figures.

The project repeatedly addresses one question: what is the best way to represent a given idea?

Transformer requires following data through a computation to see how causal masking changes attention weights. Anatomy requires rotating, looking inside, and understanding relative positions. Mechanical assemblies require taking apart parts and inspecting connections. Historical clothing requires clearly seeing overlapping panels, structure, and folds. Each kind of knowledge demands a different presentation. The author also notes that choosing the right renderers and assets during iteration is itself part of the teaching process.

In the video is a Transformer attention example. On the left, the positions of queries, keys, and values are laid out; in the middle, a similarity matrix is shown; on the right, a multiple-choice question is posed: who can a token read? Can strong similarity cross the mask? The page has 'Save prediction' and 'Try a data flow,' letting learners first save a judgment and then change the data flow to see the result.

The interaction flow can be summarized as: predict, try, observe, explain, then apply to a new situation. The author gives several questions: Does adding one more gear always make the output slower? If the radius of a conic section doubles, what happens to the area? Will a rocket take off as long as the engine produces thrust? These questions all require learners to make a prediction first, then change parameters, and finally reach a result they can reason about.

Learning science was also built into the design. Dual coding lets textual explanations and visual elements work together. Spatial and temporal contiguity determines where explanations are placed and how they stay synchronized with changes in the visuals. Cognitive load determines what is shown first and what is shown later.

The current demo covers Transformer attention, chest anatomy, rocket structure and flight, conic sections, Ming dynasty robes, and gear ratios. It supports Chinese and English and can run offline.

A comment mentioned that agents can generate decent slides but struggle with the sequence and pacing of a lesson. The author replied that the hard part of learning is that once you have learned something, you cannot unlearn it; it is not easy to reproduce your initial psychological and cognitive state. So capturing your own learning problems and turning them into a product may be easier to scale.

Another comment asked whether the numbers in the figures are computed in real time or hand-drawn approximations. Jerry Liu mentioned that when making animated explanations of Transformer internals, every number on screen comes from real computation, such as actual attention weights and cache sizes, because that is how trust is built. The author replied that they are working on another skill to visualize application repositories and LLM repositories to help understand architecture and data flow. If you need to modify a workflow or component, it will also generate a refactoring handoff document for the coding agent.

Under the gear demo, someone also noticed a detail: leaving both sets of results on screen at the same time, with the same half-turn rotation in different directions, makes the comparison immediately understandable. That is more useful than just watching an animation.

Original post: https://x.com/yiyibuilds/status/2116289366784168399