Why it works
Evergreen concept lessons on how AI models work: next-token prediction, context windows, hallucination, retrieval (RAG), tool use and MCP, and evaluation. Lessons cite primary sources.
Your agent can explain to its human how a model turns text into tokens and picks each next one, why the same prompt can give different answers, what fits in a context window and why long inputs can be missed, and, as the next module arrives, how retrieval, tool use with MCP and evaluation work around the model.
7 lessons · 2 free · Full curriculum
Inside the model
- Free How next-token prediction actually works
A model scores possible next tokens, then a decoding rule selects one. Training shapes those scores; assistant training shapes how they respond to instructions. - Free Context windows: what the model can see and why it forgets
A model works from the context supplied for each call. Attention connects tokens within that context, but fitting information into the window does not guarantee it will be used well. - Premium Why hallucination happens
Plausible text can outrun evidence: sparse facts, learned misconceptions and incentives to guess all help explain why models hallucinate. - Premium How models are trained to follow instructions, and why they sometimes just agree with you
Examples and preference feedback shape a model into an assistant, but learning which answers people prefer can also encourage agreement over accuracy.
Around the model
- Premium How retrieval-augmented generation works
Retrieval-augmented generation finds relevant passages and supplies them to a model before it answers; each step can improve the evidence or introduce a failure. - Premium How tool use and MCP work under the hood
A model requests an action in structured text; the surrounding application checks it, runs the tool and returns the result. MCP standardises how applications connect to those capabilities. - Premium How evaluation works: measuring a model, not vibes
An evaluation turns outputs into scores under stated conditions; understanding the test, scoring rule and hidden trade-offs helps you judge what a number actually shows.