My weekly assignment work for YZ50 — a 12-week intensive program that takes participants in Turkey and trains them to build AI models from scratch rather than call them through an API.
The curriculum follows Andrej Karpathy's Zero to Hero path: a single neuron written without libraries, then autograd, PyTorch, character-level language models, embeddings and backpropagation by hand, ending in a small GPT built from the ground up. Mentored by researchers from Freya (YC S25).The program's stated goal is producing AI models in Turkey, not just users of them.
| Week | Topic | Files |
|---|---|---|
| 1 | Single neuron → layer → loss → manual parameter search → numerical gradient descent | w1/ |
| 2 | Backpropagation: a small micrograd — Value autograd engine, backward(), and an MLP trained with gradient descent |
w2/ |
| 3 | Bigram character language model, built by counting and again as a single-layer network trained with gradient descent; negative log likelihood, smoothing, and the same two models rerun on Turkish names | w3/ |
| 4 | MLP language model over a three-character context: learned embeddings, minibatch training and a learning rate sweep, then a look inside the network — high initial loss, tanh saturation, Kaiming init and BatchNorm — and the same model on Turkish names | w4/ |
Each week's folder has its own README with the assignment list.