Lab 1 — Your First AI API Call

Session 1 · Generative AI: Foundations and Applications · TCE

Runs entirely in your browser (Google Colab). Any laptop or lab PC works — nothing to install, no GPU, no payment. See ZERO-SETUP. 50 minutes · pairs (both partners run everything) · Goal: your code talks to a frontier AI model

Pairs: both partners create a key. One drives per part; swap at each checkpoint.

AI coding assistants (Copilot, ChatGPT, Gemini) are allowed and encouraged in every lab. One rule: your eval set judges their code too.

---

Part A — Confirm access and get your API key (10 min)

If you completed the pre-class access check, continue with your own key. If not, ask the instructor to confirm live access before spending time debugging account setup.

1. Go to https://aistudio.google.com → sign in with a personal Google account.
2. Click Get API key → Create API key. Copy it somewhere private.
   • Already have a Google Cloud account? AI Studio creates no default project for you: Dashboard → Projects → Import your existing project first, then Get API key.
   • If the notebook’s Cell 2b (models.list()) fails, create a new key — Google rejects old “Standard” keys since Sep 2026.
3. Rules: your key = your password. Never paste it in code cells, never WhatsApp it, never commit it to GitHub.
4. While you are signed in: put two photos (a receipt, a page of your handwriting) in your Google Drive now — Lab 3 needs them.

Stuck after 5 minutes (college account blocked, phone verification loop)? Stop and tell the instructor. Every notebook has an offline path — set MOCK = True in Cell 2 only when the instructor says so (canned answers, no API) — or follow the instructor-controlled demonstration; do not share keys or create extra accounts to bypass the problem.

Part B — First call (10 min)

1. Open the Colab notebook: link on the board (session_1_lab.ipynb).
2. File → Save a copy in Drive (so your work persists).
3. Run Cell 1 (installs the SDK) and Cell 2 (asks for your key via a hidden password box; also defines the ask() helper that retries on a 429 rate limit).
4. Run Cell 2b — lists the models your key can see. If it errors, go back to Part A and make a new key.
5. Run Cell 3 — your first API call.

✓ Checkpoint 1: the notebook prints a response written by Gemini. Read it out loud to your partner.

Part C — The five prompts (10 min)

Cell 4 runs all five prompts in one loop (one ask() call each). Read every output, then write one line per prompt in the notebook’s observation cell: what was good, what was off?

#	Skill	Prompt
1	Explain	“Explain how UPI works to a 10-year-old, in 5 sentences.”
2	Summarize	“Summarize the plot of Ponniyin Selvan in exactly 3 bullet points.”
3	Translate	“Translate to formal Tamil: ‘The exam has been postponed to next Monday.’”
4	Extract	“Extract name, degree, year as JSON from: ‘Hi, I’m Priya, third year BE CSE at TCE.’”
5	Roleplay	“You are a strict interviewer at a product company. Ask me one DSA question, wait for my answer.”

✓ Checkpoint 2: all 5 ran; you wrote 5 one-line observations.

Part D — One prompt, three models (10 min — this is the part to cut if the session is running late)

Pick one prompt from Part C (or write your own). Run the same text on:

1. Gemini (your notebook — Cell 5)
2. ChatGPT — https://chatgpt.com (sign-in may be required)
3. Any third: Claude (https://claude.ai), Copilot, Meta AI — your choice

Blocked on the college network or no account? Use your phone, or compare your default gemini-flash-lite-latest against gemini-flash-latest via ask(prompt, model=...) — same lesson, no login. The flash alias now thinks before it answers, so time both: the same prompt should come back slower and more verbose from flash, and that gap is the lesson.

Fill this table (also in the notebook):

	Gemini	ChatGPT	Third model
Length / format
Tone / personality
Accuracy issues?
Which would you ship to a user?

✓ Checkpoint 3 (show instructor): your printed Gemini response + the one difference that surprised you most.

Part E — Build your test set (last 10 min, everyone)

Part E is the hand-off to Lab 2 — protect it. Sessions run back-to-back, so do this now, before the break, even if Part D was cut:

Pick one subject you know cold — DSA, cricket stats, Tamil cinema, Ilaiyaraaja discography, your hometown, anything. In a text file, write 10 questions, each with a one-word or one-number answer — a name, a year, a score. Not a sentence. Make them details only a fan would know — deep cuts, not headlines. Keep it open — this is your ammunition for Session 2’s lie-detector lab, where you’ll measure how often the AI gets your subject right.

Stretch goals (if you finish early)

• S1 · Temperature: in the notebook’s temperature cell, run the same prompt at temperature=0.0 three times, then temperature=1.5 three times. What changed? (temperature is deprecated on Gemini 3.x but still accepted — if the two runs look identical, that is the finding.)
• S2 · Tamil stress test: ask the same factual question in English and Tamil. Same answer? Same quality?
• S3 · Token counting: use the token-count cell to measure an English sentence vs its Tamil translation. Confirm what you saw in the lecture.
• S4 · The 15-line language model: run the counting-based next-word model from the lecture — train it by tallying, then sample sentences. Same three steps as Gemini; step 2 is a table you can print instead of a learned function.
• S5 · Break it: find a question Gemini answers wrongly with confidence. Your first hallucination catch — screenshot it for bragging rights in Session 2.

Overnight (before Day 2 — 5 min, important)

Day 2 starts with “chat with your own notes” (RAG). Put 2–3 documents you actually care about on your laptop or Drive: lecture notes, a textbook chapter PDF, your own study notes. Your Day 2 apps are only as interesting as the documents you bring.