25-Day Programme · 30 Minutes Per Day
WEEK 2 OF 5 · SESSIONS 6–10 · WORK AT YOUR OWN PACE
Domain 1 · Generative AIDAY 6
How Generative AI Models Actually Work
Why This Matters
Welcome to Week 2. Last week you learned what AI can do. This week you learn what is happening inside the machine — and today is the big one. When you can explain how a generative model turns your sentence into a useful answer, you stop being someone who is merely impressed by AI and become someone who understands it. That understanding is exactly what separates a candidate an employer trusts from one they do not. Read slowly today, and do not worry if it takes two passes. Every serious person in this field had to learn these words for the first time once.
30-Minute Module
0:00 – 5:00
Tokens: how a model reads
- A model does not see words the way you do. It breaks your text into small pieces called tokens. A token is roughly a word or part of a word — “Lagos” might be one token, “unbelievable” might be three (“un”, “believ”, “able”).
- Everything you send in and everything the model sends back is counted in tokens. This matters for real work because you are billed per token, and every model has a limit on how many tokens it can handle at once.
- Rough rule to remember: about 750 words is roughly 1,000 tokens. You do not need to count them by hand — you just need to know the unit exists and why it costs money.
- When I put a budget alert on your resource group last week, tokens are ultimately what that budget is measuring. Every token is a fraction of a naira of your subscription.
When someone says a model has a “128k context window”, they mean it can hold 128,000 tokens in view at once. Context window = memory span, measured in tokens.
5:00 – 15:00
Embeddings and transformers: how a model understands
- Once your text is in tokens, the model turns each token into a long list of numbers called an embedding. Think of an embedding as coordinates on a giant map of meaning: words with similar meanings sit close together. “King” and “queen” land near each other; “banana” is far away.
- This is the trick that lets a computer work with meaning at all. It cannot do maths on the word “happy” — but it can do maths on the numbers that represent “happy”. Embeddings are how language becomes arithmetic.
- The engine that does the heavy work is called a transformer. Its key ability is attention: for every token, it weighs up which other tokens matter most. In “the market in Kano was busy, so it closed late”, attention is what lets the model know “it” means the market.
- The “GPT” in ChatGPT stands for Generative Pre-trained Transformer. You now know what all three words mean. That is not trivia — that is the architecture of the whole field.
You do not need to build a transformer — nobody expects that at this level. You need to be able to say, in one sentence, that it uses attention to decide which words matter to each other.
15:00 – 25:00
Inference and temperature: how a model answers
- When you send a prompt and the model produces a response, that act is called inference. Training is when the model learns (done once, by Microsoft, on enormous machines). Inference is when it is used (done every time by you). You will only ever do inference, and inference is what you pay for.
- Here is what actually happens: the model predicts the next token, one at a time. It looks at everything so far and asks “what token most likely comes next?”, adds it, then repeats. A whole paragraph is built one token at a time.
- Temperature is the dial that controls how adventurous those choices are. A low temperature (near 0) makes the model pick the safest, most predictable next token every time — good for facts, code, and anything where you want the same answer twice. A high temperature (near 1) lets it take chances — good for brainstorming, stories, and variety.
- Real example: for a customer-service bot answering “what are your opening hours?” you want low temperature — the same correct answer every time. For a tool suggesting business-name ideas, you want high temperature.
If a model keeps giving you slightly different answers to the same factual question, the temperature is probably too high. This is a real setting you will adjust in Foundry in Week 3.
25:00 – 30:00
Fix it in your own words
- Out loud, trace one sentence through the machine: your words become tokens, tokens become embeddings (numbers on a map of meaning), a transformer uses attention to weigh them, and inference produces the answer one token at a time, with temperature deciding how bold each choice is.
- Write those six words down the side of a page in your AI-901 Notes: tokens, embeddings, transformer, attention, inference, temperature. If you can define each in one line, you have today.
- Then do the assignment. This is a thinking day, not a typing day — the value is in explaining it back.
Do not move on until those six words feel like yours. Everything in Weeks 3 to 5 sits on top of this one lesson.
Key Terms
Token
The small piece — a word or part of a word — that a model breaks text into. You are billed per token, and every model has a token limit.
Embedding
A list of numbers representing a token’s meaning, so that words with similar meanings sit close together. How language becomes maths.
Inference
The act of using a trained model to produce a response to your prompt. Training happens once; inference happens every time you use it.
Temperature
A setting from about 0 to 1 controlling randomness. Low = predictable and consistent; high = creative and varied.
Today’s Assignment
Explain how a generative AI model turns a question into an answer — in your own words, in plain language, as if teaching a friend.
- Roughly 250–350 words. Use all six terms correctly at least once: token, embedding, transformer, attention, inference, temperature. No copying, no asking a chatbot to write it — I am marking your understanding.
- Include one worked example of your own: describe a task where you would set temperature low, and one where you would set it high, and say why.
- Finish with one sentence: which of the six ideas was hardest to grasp, and how you finally made sense of it.
Submit assignment 6 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. You ask a model the same factual question three times and get three slightly different answers. Which setting would you most likely lower?
- A. The number of tokens
- B. The temperature
- C. The embedding
- D. The attention
B. The temperature — Temperature controls randomness. A high value makes the model take chances, so answers vary. Lowering it toward 0 makes it pick the most predictable next token each time, which is what you want for consistent, factual responses.
2. What is an embedding?
- A. The building where servers are kept
- B. A list of numbers that represents the meaning of a token
- C. The price you pay per request
- D. A type of AI model that only handles images
B. A list of numbers that represents the meaning of a token — Embeddings turn language into numbers so that words with similar meanings sit close together. This is what lets a computer do “maths on meaning” — the foundation of how these models work.
3. In AI, what does ‘inference’ mean?
- A. Teaching the model from scratch on huge datasets
- B. Using a trained model to produce a response to a prompt
- C. Deleting a resource group
- D. Breaking text into tokens
B. Using a trained model to produce a response to a prompt — Training is the one-time learning phase done by the model’s creators. Inference is every-day use — and it is what you are billed for each time you send a prompt.
Exam Objectives Covered
TokensEmbeddingsTransformers & attentionInferenceTemperature & parametersHow generative models work
Domain 1 · Responsible AIDAY 7
Responsible AI I: Fairness, Reliability & Safety
Why This Matters
Here is something most beginners miss, and it will make you stand out: Microsoft does not just test whether you can build AI — it tests whether you can build it responsibly. There are six principles of Responsible AI, and they appear on the exam every single time. But this is not box-ticking. When you understand these, you become the kind of person a company can trust with real customers and real data. Today we take the first three — fairness, reliability, and safety. Learn them properly and you will already sound more professional than people twice your experience.
30-Minute Module
0:00 – 5:00
The six principles — the whole map first
- Microsoft’s Responsible AI framework has six principles. Learn the list today, because the exam loves to give you a scenario and ask “which principle is this?”
- They are: Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, and Accountability.
- Today covers the first two clusters — fairness, and reliability & safety. Tomorrow covers the other four. Split like this, they are easy to hold in your head.
- A memory hook: think “FRPITA” if it helps — but understanding beats memorising, because the exam describes situations, it does not just ask for the list.
Write the six principles at the top of a fresh page now. You will fill two in today and four tomorrow. By Thursday you will know them cold.
5:00 – 16:00
Fairness: treating people equally
- Fairness means an AI system should treat all groups of people equally and not disadvantage anyone because of things like gender, age, or where they come from.
- How does unfairness even get in? Through the training data. If a hiring model learned from ten years of records where mostly men were hired, it can quietly learn to prefer men — not because anyone intended it, but because the data carried that pattern. This is called bias.
- Make it real: imagine a loan-approval AI trained mostly on data from one city. It might unfairly reject good applicants from a region it saw little of — people who look a lot like you and me. That is a fairness failure, and it is exactly the kind of harm this principle exists to catch.
- What you do about it: check that your data represents everyone the system will serve, and test the model’s results across different groups to make sure one group is not being treated worse.
If an exam scenario mentions a system giving worse results to one group of people — women, older applicants, one ethnic group — the principle is almost always Fairness.
16:00 – 26:00
Reliability & Safety: working correctly, even when things go wrong
- Reliability & Safety means an AI system should perform consistently and safely, including when it meets situations its makers did not expect.
- Why it matters: AI now helps drive cars, read medical scans, and approve payments. A system that works “usually” is not good enough when the failure could hurt someone. It must fail safely — for example, handing control back to a human rather than guessing.
- How you build it in: test the system hard before release, including with strange and unexpected inputs; monitor it after release; and design it so that when it is unsure, it says so or asks a person, rather than confidently making something up.
- Connect it to Day 6: models predict the most likely next token — they can produce a confident-sounding answer that is simply wrong. That behaviour is called a hallucination, and guarding against it is part of reliability and safety.
Reliability & Safety is the “does it keep working, and does it fail gracefully?” principle. Think self-driving car handing back to the driver when the road confuses it.
26:00 – 30:00
Lock it in
- Say the six principles out loud in order. Then define today’s two in one sentence each without looking.
- For each of today’s two, write down one real-world example in your AI-901 Notes — ideally one from Nigeria, because the exam rewards people who can connect the principle to a real situation.
- Do the assignment. It asks you to spot the principle in a scenario — which is exactly how the exam will test you.
These principles are not just exam points. In an interview, “I always think about fairness and how a model might be biased by its training data” is the sentence that gets you taken seriously.
Key Terms
Responsible AI
Microsoft’s framework of six principles for building AI that is ethical, safe and trustworthy. It appears on every AI-901 exam.
Fairness
The principle that an AI system should treat all groups of people equally and not disadvantage anyone unfairly.
Bias
When a model learns an unfair pattern from its training data — for example favouring one group — usually without anyone intending it.
Reliability & Safety
The principle that an AI system should work consistently and fail safely, even in situations its makers did not anticipate.
Today’s Assignment
Match the Responsible AI principle to the situation, then reason about a real harm.
- Part 1. For each scenario, name whether it is a Fairness problem or a Reliability & Safety problem, and write one sentence saying why: (a) A CV-screening AI rejects almost all applicants from one region. (b) A medical AI gives a confident diagnosis on a blurry scan instead of flagging that the image is unclear. (c) A voice assistant understands men’s voices far better than women’s. (d) A self-driving system speeds up when it meets a road layout it has never seen.
- Part 2. List all six Responsible AI principles from memory (you may check afterwards, but try first).
- Part 3. Think of one AI system that could affect people in Nigeria — a bank, a hospital, a government service — and describe in one short paragraph how bias in the training data could make it unfair, and what you would check to catch it.
Submit assignment 7 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. A loan-approval AI approves men far more often than equally-qualified women, because it learned from historical data. Which Responsible AI principle does this violate?
- A. Reliability & Safety
- B. Transparency
- C. Fairness
- D. Privacy & Security
C. Fairness — The system is disadvantaging one group of people based on gender, a pattern it absorbed from biased training data. That is the textbook definition of a fairness failure.
2. What is the best description of an AI ‘hallucination’?
- A. The model refuses to answer
- B. The model produces a confident answer that is actually false
- C. The model runs out of tokens
- D. The model is trained on images
B. The model produces a confident answer that is actually false — Because a model predicts the most likely next token, it can sound certain while being wrong. Guarding against this is part of the Reliability & Safety principle.
3. How many principles are in Microsoft’s Responsible AI framework?
- A. Three
- B. Four
- C. Six
- D. Ten
C. Six — Fairness; Reliability & Safety; Privacy & Security; Inclusiveness; Transparency; and Accountability. This is a guaranteed exam fact — know the list and know what each one means.
Exam Objectives Covered
Responsible AIFairnessBias in dataReliability & SafetyHallucination
Domain 1 · Responsible AIDAY 8
Responsible AI II: Privacy, Inclusiveness, Transparency & Accountability
Why This Matters
Yesterday you learned two principles. Today you complete the set — and by the end of these 30 minutes you will hold all six Responsible AI principles in your head, ready for any scenario the exam throws at you. These last four are the ones that show real professional maturity: keeping people’s data safe, making sure the system works for everyone including people with disabilities, being honest about how it works, and making sure a human stays responsible. This is the difference between a technician and someone a company puts in front of clients.
30-Minute Module
0:00 – 8:00
Privacy & Security: protecting people’s data
- Privacy & Security means an AI system must protect the personal data it uses and keep it secure from people who should not see it.
- AI is hungry for data, and a lot of that data is personal — names, faces, health records, messages. The principle says: collect only what you need, keep it safe, and respect the rules about how it can be used.
- Connect it to Week 1: remember the API key I warned you never to share? That is Privacy & Security in practice. A leaked key can expose data and authorise spending. Protecting keys and data is the same principle.
- Make it real: if you built a health app that reads patient records, you would need to make sure those records are encrypted, that only authorised people can see them, and that the AI provider is not quietly keeping copies it should not.
Privacy is about the data: whose is it, who can see it, is it kept safe? If a scenario is about leaked or misused personal information, this is your principle.
8:00 – 16:00
Inclusiveness: working for everyone
- Inclusiveness means AI should work for people of all abilities and backgrounds — including people with disabilities — and not leave anyone out.
- How it differs from Fairness: Fairness is about not disadvantaging a group; Inclusiveness is about actively designing so everyone can use it. They are close cousins, and the exam will test whether you can tell them apart.
- Make it real: speech recognition that also works for someone with a speech impairment; an app that works with a screen reader for a blind user; a service that handles many languages, not just English. That last one matters to you directly — think of the many languages spoken across Nigeria.
- Good inclusive design usually makes the product better for everyone, not only the group it was meant for. Captions help deaf users and also anyone in a noisy market.
Inclusiveness = “can everyone actually use this, including people with disabilities?” If the scenario is about accessibility or leaving people out, choose Inclusiveness.
16:00 – 25:00
Transparency and Accountability: honesty and responsibility
- Transparency means people should understand how an AI system works and how it reaches its decisions — especially decisions that affect them. If an AI declines your loan, you deserve to know it was an AI and to have some idea why.
- Transparency includes being honest about a system’s limitations — telling users what it is not good at, so they do not over-trust it.
- Accountability means the people who build and run AI stay responsible for it. The AI is never “to blame” — a human or an organisation always answers for what it does. There must be a person who owns the outcome.
- These two work as a pair: Transparency lets people see how the system behaves; Accountability makes sure someone answers for it. Together they keep AI under human control.
Easy way to separate them: Transparency = “can we understand it?” Accountability = “who is responsible for it?” Understanding versus responsibility.
25:00 – 30:00
Lock in all six
- Now say all six aloud with a one-line meaning for each: Fairness (treat groups equally), Reliability & Safety (work consistently, fail safely), Privacy & Security (protect data), Inclusiveness (work for everyone), Transparency (be understandable and honest), Accountability (humans stay responsible).
- Test yourself with the trickiest pair: Fairness vs Inclusiveness, and Transparency vs Accountability. If you can explain the difference in each pair, you have mastered this topic.
- Do the assignment — it deliberately mixes all six so you practise choosing under pressure, just like the exam.
You are eight days in and you can now name and explain the entire Responsible AI framework. Most working adults cannot. Keep that momentum.
Key Terms
Privacy & Security
Protecting the personal data an AI uses and keeping it secure. Guarding your API key is this principle in daily practice.
Inclusiveness
Designing AI so people of all abilities and backgrounds — including those with disabilities — can use it.
Transparency
Making it possible to understand how an AI works and decides, and being honest about its limitations.
Accountability
The principle that people, not the AI, remain responsible for what an AI system does. A human always owns the outcome.
Today’s Assignment
Sort six scenarios into the six principles — then argue one hard case.
- Part 1. Match each scenario to one of the six principles and give a one-sentence reason: (a) An app stores users’ medical records without encrypting them. (b) A company cannot explain why its AI denied someone a job, and no one will take responsibility. (c) A voice assistant does not work for users who are deaf or have speech difficulties. (d) A bank’s AI declines loans and never tells applicants a reason. (e) An AI performs worse for one ethnic group because of skewed training data. (f) A medical AI keeps operating confidently even when its sensors clearly fail.
- Part 2. Pick the pair you find easiest to confuse — Fairness vs Inclusiveness, or Transparency vs Accountability — and explain the difference in your own words, with an example of each.
- Part 3. In one paragraph: if you were building an AI service for people across Nigeria, which two principles would you worry about most, and why?
Submit assignment 8 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. A bank’s AI rejects a loan application and the applicant is given no explanation and cannot find out why. Which principle is most directly at stake?
- A. Inclusiveness
- B. Transparency
- C. Reliability & Safety
- D. Privacy & Security
B. Transparency — People affected by an AI decision should be able to understand how and why it was made. Being kept completely in the dark about a decision that affects you is a transparency failure.
2. What best distinguishes Inclusiveness from Fairness?
- A. They are exactly the same principle
- B. Fairness is about not disadvantaging a group; Inclusiveness is about designing so everyone, including people with disabilities, can use it
- C. Fairness applies to images; Inclusiveness applies to text
- D. Inclusiveness is only about pricing
B. Fairness is about not disadvantaging a group; Inclusiveness is about designing so everyone, including people with disabilities, can use it — They are close, and the exam uses that. Fairness = equal treatment; Inclusiveness = accessible design for all abilities and backgrounds.
3. “A human or organisation must always remain responsible for what the AI does.” Which principle is this?
- A. Accountability
- B. Transparency
- C. Fairness
- D. Inclusiveness
A. Accountability — The AI is never the one “to blame”. Accountability means a person or organisation owns the outcome. Pair it with Transparency: one is about understanding the system, the other about who answers for it.
Exam Objectives Covered
Privacy & SecurityInclusivenessTransparencyAccountabilityAll six RAI principles
Domain 1 · Choosing ModelsDAY 9
Choosing the Right Model: Capabilities, Deployment & Configuration
Why This Matters
You would not use a lorry to deliver a single letter, or a bicycle to move a tonne of cement. Choosing an AI model is the same kind of judgement — and it is a judgement employers pay for. Today you learn how to look at a task and pick a model that fits: the right capability, the right size, the right cost. This is the bridge between the concepts of Weeks 1 and 2 and the hands-on Foundry work that starts on Monday. Get comfortable here and next week’s practical work will feel natural.
30-Minute Module
0:00 – 8:00
Capabilities: match the model to the job
- Different models are good at different things. Choosing well starts with asking: what does this task actually need?
- Large language models handle text — writing, summarising, answering, reasoning. Multimodal models can also handle images and sometimes audio, so they can “look” at a picture and describe it. Embedding models exist just to turn text into those number-lists from Day 6, for search and comparison. Image-generation models create pictures from a description.
- So the first question is capability: if the task is “read a photo of a receipt”, you need a model that can see — a multimodal one. If it is “summarise this report”, a text model is enough and cheaper.
- In Foundry, all of these live in the model catalog — a menu of models you can browse and deploy. You will open it for real next week.
Always start from the task, not the model. “What does the job need the AI to do?” comes before “which model is newest?”
8:00 – 16:00
Bigger is not always better: capability vs cost vs speed
- Models come in different sizes. Bigger models are generally more capable but cost more per token and can be slower to respond. Smaller models are cheaper and faster but less powerful.
- This is a real trade-off you will make on the job. For a simple task — sorting messages into “complaint” or “not complaint” — a small, cheap model is the professional choice. For complex reasoning across a long document, a larger model earns its cost.
- Tie it back to your own account: every token you spend is real money against your free credit. Choosing a right-sized model is not stinginess — it is exactly the cost-awareness a company wants in someone they trust with a cloud budget.
- Other things to weigh: how fast the answer must come back (latency), how many requests per minute you are allowed (rate limits), and whether the task needs images or just text.
In an interview: “I’d pick the smallest model that does the job well, to control cost and latency.” That single sentence signals real cloud maturity.
16:00 – 25:00
Deployment and configuration parameters
- Choosing a model is not enough — you must deploy it. A deployment is your own running copy of a model, inside your resource group, that your code can send prompts to. You will create your first one on Day 13.
- Once deployed, you tune how it behaves with configuration parameters. You already know the most important one: temperature (Day 6) — low for consistent, high for creative.
- Two more to know by name: max tokens caps how long the response can be (which controls cost and cuts off runaway answers); and Top-P is another randomness control, like a cousin of temperature — you usually adjust one or the other, not both.
- You do not need deep maths on Top-P for AI-901. You need to recognise it as a parameter that, like temperature, influences how varied the output is.
Remember the shape: choose a model → deploy it into your resource group → tune it with parameters like temperature and max tokens. That is the workflow of Week 3.
25:00 – 30:00
Fix it in your own words
- Out loud, answer: what three things would you weigh when choosing a model? (Capability for the task, cost, and speed — plus whether it needs to handle images.)
- Note the three parameters in your AI-901 Notes: temperature, max tokens, Top-P — with a one-line meaning for each.
- Do the assignment. It asks you to choose and justify, which is the real skill here — not memorising, but reasoning.
Tomorrow is a consolidation day — a drill across everything in Weeks 1 and 2. Come in ready to be tested, and you will see how much you already know.
Key Terms
Model Catalog
The menu of available models in Microsoft Foundry that you can browse and deploy, from small text models to multimodal and image-generation models.
Multimodal Model
A model that handles more than one kind of input — for example text and images together — so it can “see” a picture and describe it.
Deployment
Your own running copy of a chosen model, created inside your resource group, that your code sends prompts to.
Max Tokens
A parameter capping how long a response can be. It limits cost and stops runaway answers. Set alongside temperature and Top-P.
Today’s Assignment
Play the role of a solution designer: choose a model for three tasks and justify each choice.
- For each task, say (i) what capability is needed — text only, or multimodal that can see images; (ii) whether you would lean toward a smaller/cheaper or larger/more capable model; and (iii) one sentence of reasoning: (1) Automatically sort 10,000 incoming customer messages into “complaint” or “not complaint”. (2) Read photos of handwritten delivery notes and pull out the address. (3) Write detailed, well-reasoned answers to complex legal questions.
- Then explain, in your own words, what a deployment is and where yours would live (name your resource group).
- Finally: for a customer-service bot that must give the same correct answer every time, what temperature would you choose and why?
Submit assignment 9 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. You need to read information from photographs of paper receipts. Which capability must your chosen model have?
- A. It must be an embedding-only model
- B. It must be multimodal (able to handle images, not just text)
- C. It must have the highest possible temperature
- D. It must be the largest model available
B. It must be multimodal (able to handle images, not just text) — The task involves seeing a photo, so a text-only model cannot do it. Start from what the task needs: here, the ability to process images.
2. For a very simple, high-volume task like sorting messages into two categories, which is usually the professional choice?
- A. The largest, most expensive model, to be safe
- B. A smaller, cheaper, faster model that does the job well
- C. An image-generation model
- D. It makes no difference which model you pick
B. A smaller, cheaper, faster model that does the job well — Bigger is not automatically better. For a simple task, a right-sized model controls cost and latency — exactly the judgement an employer wants from someone using their cloud budget.
3. What does the ‘max tokens’ parameter control?
- A. How creative the model is
- B. Which directory you are signed into
- C. The maximum length of the model’s response
- D. The number of people who can use the model
C. The maximum length of the model’s response — Max tokens caps how long the answer can be, which limits cost and prevents runaway output. Temperature and Top-P control randomness; max tokens controls length.
Exam Objectives Covered
Model capabilitiesModel catalogCost & latency trade-offsDeployment optionsConfiguration parameters
Domain 1 · ConsolidationDAY 10
Week 2 Consolidation: Domain 1 Drill + Python Refresher
Why This Matters
You are two weeks in and at the halfway mark of the whole programme. Today we do not add new material — we lock in what you have. Consolidation days are where real learners pull ahead, because knowledge you review is knowledge you keep. We will drill Domain 1 — nearly half your exam — and warm your Python hands back up before next week goes hands-on inside Foundry. Treat today as a checkpoint: prove to yourself how much you already carry.
30-Minute Module
0:00 – 8:00
Domain 1 rapid recall — say it before you read it
- Cover the screen and answer aloud: the five AI workloads from Week 1 (generative & agentic, text analysis, speech, computer vision, information extraction).
- Then the four text-analysis techniques: sentiment analysis, entity detection, key phrase extraction, summarisation.
- Then this week’s six words: token, embedding, transformer, attention, inference, temperature — one line of meaning each.
- Then the six Responsible AI principles: Fairness, Reliability & Safety, Privacy & Security, Inclusiveness, Transparency, Accountability.
- Anything you stumbled on, go back and re-read that day’s panel now. Stumbling is information, not failure — it tells you exactly where to spend five minutes.
Saying it from memory before checking is far more powerful than re-reading. The struggle to recall is what makes it stick.
8:00 – 18:00
Python refresher — back into Google Colab
- Open
colab.research.google.com(the free, browser-based Python you used in Week 1 — no install needed) and start a new notebook. Type this yourself, do not paste: principles = ["Fairness", "Reliability & Safety", "Privacy & Security", "Inclusiveness", "Transparency", "Accountability"]- Now loop over it and number each one:
for i, p in enumerate(principles):
print(i + 1, p) - This brings back three Week-1 skills at once: a list, a for loop, and
print. Theenumerategives you the position and the item together — here it numbers your six principles 1 to 6. - Remember lists count from zero:
principles[0]is “Fairness”. That is why we printi + 1— so a human sees 1 to 6, not 0 to 5.
If you get a red error, read the last line first — it usually names the problem. And if you are stuck after ten minutes, email me the error. Being stuck is normal; staying stuck silently is the only mistake.
18:00 – 26:00
A function that ties it together
- Recall the four-step AI pattern from Day 5: Import → Connect → Send → Use. Next week you will use it against a real deployed model. Today, practise the shape of a function so it is second nature.
- Type this yourself:
def describe(term, meaning):
return term + ": " + meaning - Then call it:
print(describe("Inference", "using a trained model to get an answer")) - Notice:
returnhands the finished sentence back to your code;printshows it to a human. Confusing those two is the single most common beginner slip — and now you have practised the difference. - Try calling
describetwo more times with other terms you learned this week. Each call reuses the same function — that is the whole point of functions: write once, use many times.
You are not memorising code. You are building the muscle memory of Import → Connect → Send → Use, so that next week’s real Foundry code feels familiar, not frightening.
26:00 – 30:00
Look back, look forward
- Open your AI-901 Notes and read the “one new thing” you wrote each day. Ten days, ten things you did not know two weeks ago. That is real progress — look at it.
- Next week everything gets hands-on: you open Microsoft Foundry, browse the model catalog, and deploy your first real model inside your own resource group. Make sure your Azure access from Week 1 still works — sign in to
portal.azure.comtoday and confirm you can see your resource group. If you cannot, tell me now so we fix it before Monday. - Do the assignment below — a mixed self-test that mirrors the real exam.
Halfway. The people who finish this are not the cleverest — they are the ones who kept showing up for 30 minutes a day. You have shown up for ten. Keep going.
Key Terms
Consolidation
Deliberately reviewing what you have learned so it moves into long-term memory. Reviewed knowledge is kept knowledge.
enumerate
A Python tool that loops over a list and gives you both the position and the item at once — handy for numbering things.
Four-Step AI Pattern
Import → Connect → Send → Use — the shape of nearly every program that talks to an AI service. You meet it for real next week.
Function
A reusable block of code you define once and call many times. return hands a value back; print shows it to a human.
Today’s Assignment
A mixed Week 2 self-test, plus a small piece of working Python — your halfway checkpoint.
- Part 1 — concepts. Answer in your own words: (a) Explain tokens, embeddings and inference in one short paragraph. (b) Name all six Responsible AI principles and give a one-line meaning for each. (c) Explain the difference between temperature and max tokens.
- Part 2 — Python. In Google Colab, create a list of the six Responsible AI principles, loop over it with
enumerate, and print each one numbered 1 to 6. Send me a screenshot of your notebook showing the output. - Part 3 — reasoning. Pick any AI system you can think of and, in one paragraph, describe how two different Responsible AI principles apply to it.
- Part 4 — readiness. Confirm in one line that you signed in to the Azure portal today and can still see your resource group, ready for Foundry on Monday.
Submit assignment 10 →
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Your work is marked against the published rubric and comes back to you as a written letter, usually within minutes. It is also saved to your progress page, so you can re-read every letter later. Do not submit until you have passed the self-check below.
Self-Check — Exam-Style Questions
1. In Python, given
principles = ["Fairness", "Reliability", "Privacy"], what does principles[0] return?- A. "Reliability"
- B. "Fairness"
- C. An error
- D. "Privacy"
B. "Fairness" — Lists are indexed from zero, so
[0] is the first item, “Fairness”. This catches almost every beginner once — better here than in the exam.2. A model gives a confident but false answer about a fact it was never reliably taught. What is this called, and which Responsible AI principle addresses guarding against it?
- A. An embedding; Inclusiveness
- B. A deployment; Fairness
- C. A hallucination; Reliability & Safety
- D. A token; Transparency
C. A hallucination; Reliability & Safety — A hallucination is a confident, wrong answer — a natural risk of predicting the next likely token. Building systems that fail safely and do not mislead falls under Reliability & Safety.
3. You must choose a model to read text from photographs of forms, on a tight budget. What matters most?
- A. Only pick the largest model regardless of cost
- B. It must be multimodal (handle images), and you balance capability against cost
- C. Set temperature to maximum
- D. It must be an embedding-only model
B. It must be multimodal (handle images), and you balance capability against cost — The task needs a model that can see images, and on a tight budget you choose the smallest such model that does the job well. Capability first, then cost and speed.
Exam Objectives Covered
Domain 1 reviewAI workloadsResponsible AIGenerative AI conceptsPython lists & loopsFunctions