← all lessons
What's new
Updated for mid-2026The course is maintained as the field moves. Here's what's changed.
July 2026
New lessons and a mastery layer
- Two new lessons: how a vector becomes the next word (logits and sampling), and a map of the 2026 model families, base, reasoning, embedding, multimodal, agentic, and more.
- Retrieval engineering is now part of the main path, right after RAG, so 'why RAG hits or misses' is core, not optional.
- New 'Explain it in 60 seconds' checkpoints after each act: pick the clearest plain-language explanation, then compare it with a model answer.
June 2026
Current for the 2026 AI stack
- Six new rooms: model routing, AI governance, realtime voice agents, computer-use agents, a retrieval-engineering lab, and copyright & data provenance.
- Deeper interpretability (sparse autoencoders and circuits), synthetic data, and advanced evals (offline vs online, red-team sets).
- Secondary topics added: open weights vs closed models, guardrails vs evals vs permissions, agent-to-agent interoperability, and batch/async inference.
June 2026
Reason about it, not just watch it
- Every concept lesson now ends with an 'in the wild' transfer: a real product scenario that connects the mechanism to where you'll actually meet it.
- A new final capability diagnostic: ten cross-act scenarios that test whether you can diagnose real AI failures, with a path back to any lesson you miss.
June 2026
Completion you can hold
- Finishing the core path now earns a dated certificate of completion and a celebratory finish, gated so it only fires when you've genuinely completed it.
- A buyer-trust FAQ and a freshness marker, so you know what the course covers and that it's maintained.