Artificial Intelligence
Build learning from nothing: gradient descent, classical ML by hand, a neural net and an autograd engine you write yourself, up to a tiny transformer. No magic — just math you can see run.
Foundations — What Learning Is
0/5Loss, the optimizer, and linear regression as a learning algorithm you code.
Features, labels, loss — the whole setup
Turning 'being wrong' into a number to minimize
The optimizer, re-derived for learning
Fit a line by minimizing loss, in code
The train/test split and why it exists
Classical ML I
0/5Logistic regression, cross-entropy, regularization, and honest evaluation.
From a line to a probability
The loss that fits classification
L1/L2 and the bias-variance tradeoff
Learning with no model at all
Precision, recall, ROC — what to actually trust
Classical ML II
0/5Trees, ensembles, boosting, SVMs, and Naive Bayes — the pre-deep toolkit.
Learning a flowchart from data
Many weak learners into one strong one
Fitting the residuals, round after round
The maximum-margin idea
Bayes' theorem turned into a classifier
Unsupervised Learning
0/5Clustering, PCA (which is just SVD), and representation.
Finding structure with no labels
Dimensionality reduction — and why it is just SVD
Why high-dimensional space is strange
Spotting the point that doesn't belong
How you present data decides what a model can learn
Neural Networks I — Build One
0/5The neuron, the forward pass, and backpropagation derived from the chain rule.
A dot product and a nonlinearity
Matrices moving activations through the net
ReLU, sigmoid, tanh — why nonlinearity is the point
The chain rule over an entire network
No framework — just arrays and calculus
Neural Networks II — Autograd
0/5Computational graphs and reverse-mode autodiff — build your own micrograd.
Every operation as a node
How gradients flow backward, automatically
Your own tiny 'micrograd'
Stochastic gradient descent in practice
Momentum, RMSProp, Adam — derived
Deep Learning
0/5Why nets fail to train, convolutions, and sequence models.
Why nets fail to train, and the fixes
Dropout, weight decay, early stopping
Weight sharing, and how machines see
Convolutions from scratch, then trained
Memory over time, and the vanishing gradient
Modern Deep Learning
0/5Embeddings, attention, and building a tiny transformer language model.
Meaning as geometry in a vector space
The mechanism, built from scratch
Attention assembled into an architecture
A small transformer you train yourself
Data, compute, and what changes with size
Frontier & Capstone
0/5How LLMs are trained, generative models, alignment, and an end-to-end capstone.
Pretraining, fine-tuning, RLHF
VAEs and diffusion — the core idea
Where models go wrong, and how we measure it
And how to keep learning it after this
A real dataset, from raw data to evaluation, all yours