| Clock | Min | Block | Say / do — key cues |
|---|---|---|---|
| 0:00 | 4 | Welcome | Today you'll talk to AI 3 ways. Rhythm: learn a little → build a little. Get everyone into the Colab tab now. |
| 0:04 | 13 | Station 1What is this? | Builder vs researcher (you drive, not build the engine) · LLM = next-word predictor · watch it choose a word · hallucination = same trick, confidently wrong. |
| 0:17 | 10 | 🧪 ColabPractical 1 | Finish "best part of summer…" 5 ways · ask a made-up book's page count. ✅ real reply + caught a made-up fact |
| 0:27 | 8 | Station 2Meet the models | Big 3 (ChatGPT/Claude/Gemini) + open vs closed · pick by 3 dials: smarts vs speed vs price, context window, small vs large. (Vehicle for the trip.) |
| 0:35 | 10 | 🧪 ColabPractical 2 | Same hard question → a flash vs a pro model; time each. ✅ compared quality + speed |
| 0:45 | 8 | Station 3Talk with code | API = the waiter (order → dish) · tokens = what you pay for, context window = counter space · dials: temperature & max tokens. |
| 0:53 | 12 | 🧪 ColabPractical 3 | Send a prompt in code · count tokens · run at temp 0 vs 1. ✅ two temps, visibly different replies |
| 1:05 | 5 | Wrap + exit | Exit ticket: "What is an LLM?" & "Name one time it might lie to you." Celebrate — they ran real code. Tease Session 2. |
| Clock | Min | Block | Say / do — key cues |
|---|---|---|---|
| 0:00 | 2 | Welcome | Recap: AI predicts text. Today we steer it, feed it our own facts, and let it act. |
| 0:02 | 14 | Station 4Prompt engineering | Prompt = the "text so far" it builds on · 5 ingredients: Role, Task, Context, Format, Constraints · power moves: few-shot & "think step by step". |
| 0:16 | 12 | 🧪 ColabPractical 4 | Weak → strong prompt · few-shot classifier · chain-of-thought on a word problem. ✅ strong beats weak + CoT fixed a wrong answer |
| 0:28 | 13 | Station 5Memory (RAG) | Problem: amnesia (cutoff + never read your notes) · loop: retrieve → stuff → answer (open-book exam) · embeddings = meaning as numbers, vector DB, chunks. |
| 0:41 | 12 | 🧪 ColabPractical 5 | Embed notes · cosine similarity (dog↔puppy > dog↔car) · tiny RAG with vs without context. ✅ right note retrieved; with-context answer is correct |
| 0:53 | 12 | Station 6AI agents | Chatbot talks; agent acts · tools / function-calling · loop: think → act → observe → repeat · ⚠️ agents do real things — human review matters. |
| 1:05 | 12 | 🧪 ColabPractical 6 | Give the model a calculator tool; watch it call it for exact math. ✅ tool result is exact |
| 1:17 | 5 | Wrap + exit | Exit ticket: "What does RAG do?" & "What makes an agent different from a chatbot?" Tease Session 3 (the finale). |
| Clock | Min | Block | Say / do — key cues |
|---|---|---|---|
| 0:00 | 4 | Welcome | Three last climbs, then you ship something. Everything so far comes together today. |
| 0:04 | 12 | Station 7Images & sound | Multimodal = extra senses (eyes + ears) · vision: show a photo, ask about it / read text · plus image generation & voice (speech↔text). |
| 0:16 | 10 | 🧪 ColabPractical 7 | Upload a photo · ask Gemini to describe it & read any text. ✅ the description matches the picture |
| 0:26 | 13 | Station 8Safety & ethics | Hallucinations (callback) → verify, cite, "say I don't know" · bias: skewed data in, skewed out · privacy: never paste secrets · don't trust blindly. |
| 0:39 | 10 | 🧪 ColabPractical 8 | Make it invent fake sources, then fact-check · compare with a "say I don't know" instruction. ✅ caught a hallucination + saw the guardrail work |
| 0:49 | 11 | Station 9Builder toolkit | AI code helpers (Cursor, Copilot, Claude Code) · the builder's method: small & real → build → test → ship → share · starter-project menu. |
| 1:00 | 12 | 🧪 ColabCapstone | Build the Study Buddy (prompt + RAG over their own notes) that quizzes them. ✅ it retrieves the right note & quizzes from it |
| 1:12 | 5 | Finale 🏆 | Nine stations done — they're builders now. Exit ticket: "Name one thing you'll build." Point them back to the roadmap to keep shipping. |