Session 01 · Understand
How LLMs work — and when they lie
3 h · live
Why can an AI sound like a senior engineer — and still be wrong?
Learning goals
- Explain how an LLM works — tokens, context, next-word prediction — using a structural model as your example, to predict when AI will help you and when it will fail you.
- Induce, detect, and correct hallucinations — including a fabricated code clause — to turn “fluency is not evidence” into an operating reflex.
- Audit an AI conversation about a building, classifying every claim as grounded, teaching, or inference, to know exactly which statements you could sign your name under.
- Engineer prompts as control interfaces — role, audience, structure, verification — to command output quality instead of hoping for it.
Topics
- 01.1Opening challenge: a complete building analyzed, diagnosed, and designed by AI in the first fifteen minutes — driven by you
- 01.2Tokens: how an AI reads your words — and your structural model
- 01.3Next-token prediction: where fluent engineering answers really come from
- 01.4Autoregressive generation: why a confident wrong start drags the whole answer with it
- 01.5The context window: what the AI can actually see — and what it silently guesses
- 01.6Temperature and sampling: tuning determinism when AI writes engineering code
- 01.7Hallucination and grounding: the trap — a fabricated code clause your judgment must catch
- 01.8Grounded vs. inferred claims: auditing an AI conversation about your building, line by line
- 01.9Prompting as a control interface: one structural question, five prompt designs, five different answers
- 01.10The Good Answer Contract: your reusable quality checklist for AI in engineering
How it’s taught
You face the activation challenge before any theory. Then the LLM Mechanics Explorer — an interactive artifact with a live tokenizer, probability bars, and a temperature slider — makes the mechanics visible. Pair labs follow: the five-prompt experiment and the hallucination clinic, where you make the AI fail on purpose and diagnose why.
You leave with
Your LLM Behavior Field Guide: your mental model, three failure modes you produced yourself, your best prompt patterns, and your Good Answer Contract — plus one audited AI conversation with every claim labeled.
Between sessions · Bring one real interaction where an AI ignored or forgot your professional standards (≈20 min). It becomes Session 2’s raw material.