Math & ML Refresher

Why this phase exists

LLMs are not magic — they are matrices multiplied by other matrices, with a clever loss on top, optimized by gradient descent. If those three sentences feel hand-wavy, this is your phase. Skip it and the rest of the roadmap will feel like memorizing a foreign language without learning the alphabet.

What you’ll be able to do at the end

• Read a research paper and not freeze when you see ∑, ∇, or KL.
• Implement linear regression, logistic regression, and a 2-layer MLP from scratch in NumPy.
• Manually derive backprop for a tiny network and verify your gradients numerically.
• Reason about loss functions, regularization, and optimizers without copy-pasting.