You've now seen the why. The remaining seven tracks are the how. Each one zooms into a specific layer of the architecture so that by the end you can read any modern model card and reason about its design choices instead of just reciting them.
- Track 2 — Tokenization. How text becomes integers. BPE, WordPiece, SentencePiece, vocabulary size, and the source of "GPT can't count letters."
- Track 3 — Embeddings & Position. Token IDs become dense vectors; positions are injected via sinusoids, learned embeddings, RoPE, or ALiBi.
- Track 4 — Attention. Q/K/V, multi-head, causal masking, KV cache, GQA, Flash Attention. The deepest track.
- Track 5 — The Block. How attention plus FFN plus norm plus residual combine into the unit cell, and what changes when you stack 80 of them.
- Track 6 — Training & Generation. Loss, schedules, mixed precision, decoding strategies, and the alignment stack (RLHF / DPO / GRPO / Constitutional AI).
- Track 7 — Model Families. GPT, LLaMA, Mistral, Claude, Gemini, Qwen, Phi, Gemma, plus state-space alternatives (Mamba, RWKV).
- Track 8 — Practical Understanding. Reading model cards, memory math, quantization, benchmarks, scaling laws in 2026, inference optimization, and how to stay current.
Each track is self-contained but builds on previous ones. If a topic feels familiar, skim the lessons and concentrate on the exercises — that's where the actual internalization happens. Let's go.