Six tracks, each a set of guided paths. Start at the rung that matches what you already know. Every path is a short sequence of interactive explainers.
Calculus, linear algebra, and probability: the tools every model is built from.
3 paths · 20 explainers · Open track →
Derivatives, the chain rule, and the optimization math behind training
Path · 5 explainersVectors, tensors, eigenvalues, and the decompositions ML relies on
Path · 8 explainersDistributions, Bayes, estimation, and information theory end to end
Loss, gradient descent, classic algorithms, evaluation, and working with data and features.
4 paths · 27 explainers · Open track →
Prediction tasks, loss functions, gradient descent, and the knobs that drive training
Path · 8 explainersTrees, SVMs, clustering, and ensemble methods beyond neural nets
Path · 8 explainersValidation, cross-validation, the bias–variance tradeoff, and pitfalls
Path · 5 explainersCleaning data, building features, embeddings, transfer learning, and dimensionality reduction
Neurons and backprop, training deep networks, the main architectures, and computer vision.
4 paths · 25 explainers · Open track →
From a single neuron to forward propagation, loss, and backprop
Path · 7 explainersOptimizers, schedules, initialization, normalization, and keeping training stable
Path · 6 explainersCNNs, recurrent nets, autoencoders, GANs, and skip connections
Path · 4 explainersPooling, detection, segmentation, and data augmentation
Text representations, language models, transformers, diffusion, RLHF, LoRA, and RAG.
4 paths · 22 explainers · Open track →
Tokenization, classic text features, and word embeddings
Path · 5 explainersClassification, seq2seq, NER, and language modeling with BERT
Path · 6 explainersFrom attention to decoder-only transformers, the architecture behind GPT
Path · 4 explainersDiffusion models, RLHF, LoRA, and retrieval-augmented generation
How LLMs are trained at scale and served fast, plus vector search, calibration, and drift in production systems.
3 paths · 16 explainers · Open track →
Mixed precision, distributed training, scaling laws, and distillation
Path · 8 explainersSampling, the KV cache, quantization, MoE, and fast inference techniques
Path · 4 explainersVector search, chunking and reranking, calibration, and data drift
Case studies that combine what the earlier tracks teach into whole systems under real limits on latency, cost, and quality.
3 case studies · Open track →
Plan GPUs, latency, and cost for a 7B chat model from first principles
Builds on KV Cache, Quantization, Continuous Batching, Decoder-Only TransformersCase study · 6 stepsSize chunking, the vector index, search, reranking, and generation for a RAG system
Builds on Retrieval-Augmented Generation, Embeddings, Chunking & Reranking for RAG, Vector Search & HNSW, K-Means Clustering, KV CacheCase study · 6 stepsRetrieval, ranking, re-ranking, feedback loops and A/B test length for a 1M-item feed
Builds on Embeddings, Vector Search & HNSW, K-Nearest Neighbors, Gradient Boosting, Model Evaluation & Metrics, Class Imbalance & Resampling, Hypothesis TestingNext case studies: building a RAG system and designing a recommendation feed.