Two AI Certs in Two Months: How I Used AI to Compress an Exam Down to What Actually Gets Tested
AI-900 in June, the III Generative AI certification in July. Here's my prep method: real past papers first, use AI to squeeze hundreds of practice questions into a single 'only-what-I-haven't-seen' review card, and judge on the spot which chapters to skip. Plus the topics this exam actually tests. A learning note.
To stretch your gaze a thousand miles,
climb one more storey up.
—— Wang Zhihuan, “On the Stork Tower” (Tang dynasty, c. 704); translation mine
I passed AI-900 in June, and in July I picked up the Generative AI certification from Taiwan’s Institute for Information Industry (III). Two certs in two months. That pace sounds aggressive, and it was — but I didn’t get there by grinding. I got there with a method: AI collaboration, past papers first, and knowing what to skip. Here it is, nothing held back.
What this cert actually is
The spec is simple:
- Computer-based, 80 single-choice questions, 90 minutes, pass at 70 (out of 100).
- Four areas: fundamentals, prompting, applied skills (all the generation tools), and ethics & law.
This is a literacy-level exam. No coding, no cloud deployment. It checks whether you understand concepts and can tell similar terms apart. So the strategy is nothing like a developer cert — you win on breadth and discrimination, not depth.
Rule one: past papers first, and don’t let your own mock questions fool you
This was my most painful lesson from AI-900: finish every real past paper before you touch a single mock question.
Plenty of people start by asking an AI to “generate 50 practice questions.” That’s a trap. AI-written mocks tend to miss the real flavor of how the exam asks things. You breeze through them, score high, and walk away with a false sense of fluency. The real papers are the signal closest to the actual test.
Rule two: use AI to compress the bank down to what you don’t know
This is the trick I most wanted to share, and it’s where AI genuinely earned its keep.
After a few hundred questions you hit a wall: redoing them all wastes time, but you’re afraid of missing the ones you haven’t seen. So here’s what I did.
I took the full text of every question I’d already answered, treated it as a comparison set, and had the AI strip out every term and topic that had already appeared — leaving only what I’d never once touched.
The result surprised me. Special terms dropped from 71 to 21. Most of the familiar chapter points got cut too. A thick stack of notes turned into a single “fill-the-gaps” card, and in the last hours before the exam that card was all I read. Best return on time I got all week.
Doing this by hand would drive you insane — cross-checking hundreds of questions. Hand it to an AI and it takes minutes. That’s the right way to use AI: not to think for you, but to do the grunt work no human wants to do.
Rule three: on the day, decide what to skip
Right at the end I did one more thing: I worked out what I could ignore.
My review card had a whole chapter on “recent developments” — packed with cutting-edge architecture names (all the new models trying to beat the Transformer, extreme quantization tricks, and so on). None of it had shown up across the hundreds of questions I’d already done. That’s a loud signal: low frequency. A literacy exam won’t test that kind of engineering detail, maybe one or two name-recognition questions at most.
So I made the call: skim that chapter for name recognition, and spend my brain on the high-frequency stuff instead. That judgment is the one thing AI can’t give you. It can organize; betting on what gets tested is on you.
The good stuff: what this exam actually tests
If you’re sitting it too, aim your effort here (all concept-level — I’m not reproducing any real questions):
- Telling function names apart: summarization / paraphrasing / expansion / translation / sentiment analysis. You get a scenario and pick the right name. Highest-frequency question type by far.
- The three voice conversions: TTS (text to speech), STT (speech to text), S2S (speech-to-speech translation). Learn to keep them straight.
- Ethics and law: liability (when AI errs, it’s not “the AI is responsible” — it falls on the designer, deployer, or user), copyright of AI-generated content, and privacy-enhancing techniques (differential privacy, homomorphic encryption, federated learning, secure multi-party computation). This chapter is the best return on study time. Read it until it’s solid.
- Prompt attack and defense: prompt injection (the broad category) contains jailbreak (breaking safety limits, like the grandma exploit); defenses include the sandwich defense and delimiters.
- The recurring big ideas: AI is “augmentation and collaboration,” not wholesale replacement; the “double-edged sword” — every application brings efficiency and a new risk at the same time.
One honest line to close
The AI did the data wrangling and the review-card compression for me. But the judgment — how to prepare, where to place my bets, what to skip — was mine. AI is an accelerator, not a stand-in. And that’s the skill I think matters most from here on: whether you can use AI to amplify your own judgment.
One in June, one in July, and in August I plan to book one more to fill out the advanced tier. If you’re on the certification road too — get the method right, and the speed will surprise you. Let’s keep climbing.