Vectors, matrices, projections and eigendecomposition, always attached to the model they show up in.
12h · 8 lessons · 1 challenge
A matrix multiply is the entire forward pass of a linear layer, and a shape error is the most common way a training run dies. This module builds linear algebra from vectors up, but every idea arrives attached to the place it appears in a model: projections as embeddings, rank as the bottleneck in a LoRA adapter, eigenvectors as the axes PCA rotates onto.
You will hand-multiply matrices before you are allowed to call NumPy, because the people who can do that are the people who can debug a shape mismatch in thirty seconds instead of an afternoon.
Twenty broken forward passes, each one dimension wrong. Fix every one in as few edits as you can. The leaderboard ranks by edits, not time.
Worth 400 XP.
