Stations 7–9 · Give AI eyes & ears → build responsibly → ship your first real thing.
Length: ~75 min · You'll need: a laptop + the Session 3 Colab notebook · This is the finale: we finish the 9-station roadmap today
The shape of the finale
Three last climbs, then you ship
Today's three stations
Station 7 — Senses: AI that can see photos and hear audio, not just read.
Station 8 — Safety: the part adults care about most — wrong, biased, not-private.
Station 9 — Toolkit & ship: the pro tools, then your first real project goes live.
Same rhythm: learn a little → build a little → prove it in Colab before we move on.
7
Station 7 · Multimodal
Beyond text: images & sound
Concept · Extra senses
"Multimodal" = more ways in and out
A "mode" is just a type of information: text is one, images another, audio another. A multimodal model works with more than one at once — show it a photo, play it a voice, and it understands.
👀 Analogy
A text-only model is a brilliant friend who can only pass notes — smart, but blindfolded with earplugs in. Multimodal gives that friend eyes and ears. Same brain — more ways to reach it.
Concept · Vision
Give it eyes: ask about a picture
You hand the model an image + a question, and it answers in words — just like it would about a paragraph.
Describe it — "What's in this photo?" (great for alt-text so blind users can follow along).
Read text inside it — "Type out this handwriting" / "what does this sign say?" (OCR).
Check a diagram — "Is this graph trending up?" / "does my triangle look labeled right?"
🔑 The core idea
Vision doesn't replace the text brain — it feeds it. The model "looks," turns the picture into an internal description, then does its normal next-word prediction on top. Your prompting skills still apply.
Concept · Making, not just reading
A paintbrush and a voice
Words → picture
Image generation
Describe a scene, get a brand-new image (DALL·E, Midjourney & friends). Name subject, style, lighting, composition for a strong result.
Sound, two halves
Voice AI
Speech-to-text (ears — transcription) and text-to-speech (mouth — TTS). Chain them with an LLM and you get a talking assistant.
⚠️ Where generators stumble
Text inside images (signs, logos) often comes out garbled, and fine details (fingers, a specific real person) can be unreliable. Treat output as a first draft, not a final.
🧪 Practical · Colab — Station 7
Give Gemini eyes
Upload a photo — a page of handwriting, a chart, or a messy desk.
Send the image + a question to Gemini and print the answer.
Ask three different questions: describe it, read any text, what looks off?
✅ Test: Gemini correctly answered a question about YOUR photo
8
Station 8 · Responsibility
Safety & ethics
Callback · Session 1
Confidently wrong: hallucinations
⚠️ The root danger
An LLM predicts likely text, not verified-true text. When it doesn't know, it doesn't stop — it invents fake books, fake quotes, wrong dates, in the exact same sure-sounding voice. That confidence is what makes it sneaky.
Builders design around it — you can't delete it:
Verify what matters — facts, numbers, quotes, code — against a real source.
Ask it to cite, and invite "say 'I don't know' if you're not sure."
Keep a human in the loop — AI drafts, a person approves.
Concept · Fairness
Bias: skewed data in, skewed answers out
A model learns language from decades of human writing — with all our stereotypes baked in. It doesn't judge any of it; it just learns "what text usually follows what," and can quietly repeat unfair patterns.
Where it shows
Skewed patterns
Assuming a "doctor" is one gender and a "nurse" another. Weaker results for names or languages rare in the data.
Why it's serious
Deciding about people
When AI helps flag applicants, moderate posts, or grade essays, hidden bias can treat real people unfairly — at scale, unnoticed.
Bias won't announce itself — it sounds as calm and confident as everything else. So you go looking for it: test with different names, genders, and languages, and check whether the tool is consistently harder on one group.
Concept · Privacy & verify
Guard what you paste — and don't trust blindly
⚠️ Never paste into an AI tool
Passwords & API keys.Financial info — card, bank, account numbers. Other people's private info — a friend's address, someone's messages. Assume your input may be stored and read by people. If you'd hesitate to post it publicly, don't paste it.
🔑 The whole station in one line
AI is a brilliant first draft, never the final authority. It can be wrong, biased, and not private — build honestly with those facts in mind instead of pretending them away. The higher the stakes, the harder you verify.
🧪 Practical · Colab — Station 8
Catch it lying, then make it honest
Ask for 3 sources + 2 quotes on an obscure but checkable topic.
Fact-check each one — which are real, which were invented?
Re-ask with "only include sources you're confident exist; say if unsure." Compare.
✅ Test: you caught one made-up "fact" + the honesty prompt changed the reply
9
Station 9 · The finish line
Your builder toolkit
Concept · Build with a co-pilot
AI code helpers: your pair programmer
In your editor
Cursor & Copilot
AI that lives inside your code editor — autocompletes lines, explains code, rewrites chunks in plain English.
In your terminal
Claude Code
An agentic helper: give it a goal, it reads your files, writes across them, and runs commands.
In a chat window
ChatGPT / Claude / Gemini
Always-available pair-programmer — paste an error, get a plan, grab a quick function.
Start here
Your first pick
A chat assistant in one window, your editor in another. That combo alone carries a whole first project.
Concept · The builder's method
How to actually ship a first project
🔑 The method
Pick something small and real. Build it with a helper. Get it working end-to-end — real input to real output. Then deploy it so a friend can open it on their phone. Done and running beats perfect and unfinished, every time.
Your first project is the whole roadmap clicking together:
Pick a model (S2), call it through an API or no-code block (S3).
Steer with a prompt (S4), feed your own notes with RAG (S5), add an agent if it needs steps (S6).
Keep your safety habits on the whole time (S8) — especially: test for hallucinations before you share.
Concept · Steal one of these
Starter-project menu
Great first pick
📚 Study-buddy quizzer
Paste your notes, get quizzed one question at a time. (Today's capstone.)
💬 Tone-fixer
Paste a rough message, pick "friendlier" or "more formal," get a polished rewrite.
📷 Explain-this-photo helper
Upload a homework photo or diagram; the AI explains it in simple steps. (Uses Station 7!)
🎭 Themed class chatbot
A bot that answers in the voice of a book character or historical figure.
🧹 Messy-text-to-table
Paste a jumble of names and dates; get back a clean, sorted table.
🗓️ Plan-my-week helper
List assignments + free time; get a realistic study schedule you can adjust.
🧪 Capstone · Colab — Station 9
Build a study-buddy from your own notes
Paste real class notes as the facts the AI must use (that's RAG, S5).
Prompt: "Ask me one question at a time about ONLY these notes, then grade my answer."
Break it: ask about something NOT in the notes — did it make something up? Tighten the prompt.
✅ Test: it quizzed you from your notes AND refused to answer off-notes
You finished the roadmap 🏆
Nine stations. You're a builder now.
AI is a next-word predictor you can call through an API, steer with prompts, feed with RAG, wire into agents, give senses, and build with safely.
That's not trivia — it's a working toolkit, and it's yours.
🎟️ Exit ticket
In one sentence: Name one thing you'll build — and the smallest first version you could ship this week.
Keep going → Revisit the 9-station roadmap anytime, and keep shipping — one finished small project beats one giant idea you never launch.
LearnAI Workshop · Session 3 — Real-World · press S for coach notes