AWS AI Practitioner Day 1: AI, ML and GenAI Explained
Here is the whole of Day 1 in one sentence: artificial intelligence contains machine learning, which contains deep learning, which is what foundation models are built from, and generative AI is what you get when you point a foundation model at producing something new. Five terms, each one inside the last — not five things sitting side by side, which is how almost every article lists them. Draw those five circles once and use them as your map: they cover most of two exam domains between them. Then read the dates below before you book anything, because one of the two AWS machine learning exams changes version in September 2026.
The course Jake nearly paid for
Jake had a tab open with a course in it and his card on the desk. Every second customer had asked him something about AI in the past month — whether it would do their invoices, whether it was listening to them, whether he could set it up — and he had run out of ways to say he did not really know.
"It's four hundred dollars," he said. "Says it prepares you for the AWS machine learning certification. Bit steep, but I'm tired of shrugging at people."
"Which certification does it name?"
He read it out. It was the Machine Learning Specialty — an exam that has since been retired and split into two replacements.
"Then whatever else that course is, nobody has updated it in a long while. And it's aimed at engineers building models, which isn't the question your customers are asking you anyway. What you want is the foundational one, it's a hundred dollars, and Day 1 of it is about six ideas."
"Six ideas for four hundred dollars saved. Go on then."
So this is those ideas. And because Jake nearly bought preparation for an exam that no longer exists, the section on which exam comes with dates — because one of the two live ones is about to change version too.
The five circles, each one inside the last
Nearly every explainer gives you these five as a list, which quietly suggests they are alternatives. They are not. They are nested, like Russian dolls, and getting that one relationship right is worth more than memorizing thirty definitions.
| Level | The one-line version | Something you have met |
|---|---|---|
| 1. Artificial intelligence | Machines doing things that normally need human thinking | A chess program from the 1990s |
| 2. Machine learning | Learning the rules from examples instead of being given them | Your spam filter |
| 3. Deep learning | Machine learning using many-layered neural networks | Face unlock on your phone |
| 4. Foundation models | One enormous model, trained broadly, reusable for many tasks | The model behind a chat assistant |
| 5. Generative AI | Using such a model to produce something rather than label it | Anything that writes you a paragraph |
1. Artificial intelligence — the outermost circle
The broadest term, and the oldest. It covers any attempt to get a machine to do something that would otherwise need a person thinking about it: playing a game, planning a route, recognizing a face, holding a conversation.
Crucially, it does not require learning at all. A system built entirely from rules a human wrote — if the temperature goes above this, do that — is artificial intelligence in the traditional sense. A great deal of what was called AI for forty years worked exactly that way, and some of it still does, because for a well-understood problem a clear rule beats a learned guess.
2. Machine learning — where the change happened
Machine learning is the part of AI where you stop writing the rules and start supplying examples.
Take spam. The rule-based approach is to write conditions: if the subject shouts in capitals, if it mentions a lottery, if the sender is unknown. You would spend years on that list, and spammers would spend an afternoon getting around it.
The machine learning approach is to hand the system a very large pile of messages, each one marked as spam or not spam, and let it work out for itself which combinations of features predict which label. Nobody writes the rules down. Nobody can necessarily read them afterwards either, which is a genuine trade-off rather than a detail.
That is the whole idea, and Jake's version of it landed properly: "So instead of telling it what a fake invoice looks like, I show it four hundred real ones and four hundred fakes." Yes. That is machine learning.
3. Deep learning — a technique, not a level of effort
Deep learning is machine learning done with neural networks that have many layers. The "deep" refers to the number of layers, nothing more — it is not a claim about depth of understanding, which is a misreading the name invites.
A neural network is a large stack of very simple mathematical units, each passing a number to the next. Early layers pick up crude features, later layers combine them into something meaningful. In an image model the first layers find edges, the middle ones find shapes, the last ones find faces.
Two things made it work in practice: enough data, and enough computing power. The ideas are decades old. The conditions for them to succeed arrived comparatively recently, which is why the field appeared to explode from nowhere when it had actually been walking for forty years.
4. Foundation models — the shift that changed the economics
Everything above described building a model for a job. Want to detect spam, train a spam model. Want to recognize invoices, train an invoice model. Each one needs its own labeled data and its own training run.
A foundation model breaks that pattern. It is trained once, on an enormous and deliberately broad range of data, at a cost only large organizations can carry. And because it has absorbed general patterns rather than one narrow task, the same model can be pointed at summarizing, translating, classifying, answering or writing — without being retrained from scratch.
That is the change worth understanding, because it is why the last few years happened. The expensive part moved from "every company trains its own model" to "a few organizations train enormous models and everybody else adapts them." The AI Practitioner exam devotes its largest single domain to exactly this, which tells you how central it is.
5. Generative AI — a use, not a separate technology
Generative AI means using a model to produce new content — text, images, audio, code — rather than to produce a label or a number.
It is worth being precise, because this is where the vocabulary usually collapses. Generative AI is not a rival to machine learning, and it is not a newer replacement for it. It is a use of models that are themselves built with deep learning. A model that reads a review and says "positive" is not generative. A model that reads the same review and writes a reply is.
Most of what people mean when they say "AI" in 2026 is this fifth circle, which is a very small part of the first one.
And the sixth circle that arrived recently: agents
Worth adding, because it is now on the engineer-level exam. An agent is a model given the ability to take actions and use tools — look something up, call a service, run a step, check the result and decide what to do next — rather than only producing a block of text and stopping.
The model is the same. What changes is that it sits inside a loop with permission to do things. That is a meaningful difference in what can go right and, more to the point, in what can go wrong, which is why responsible-AI material has grown in step with it.
The three ways a machine learns
Exam material returns to this constantly, and it is genuinely useful outside an exam, because the three answer different shapes of question.
| Type | What you give it | In Jake's shop |
|---|---|---|
| Supervised | Examples with the right answers attached | 800 invoices marked genuine or fake → flag the next one |
| Unsupervised | Examples with no answers — find the structure yourself | Two years of sales → "these customers behave alike" |
| Reinforcement | A goal, and a score after each attempt | Try a repair order, see how long the day took, adjust |
The practical difference is who does the labeling, and it is almost always the expensive part. Supervised learning needs somebody to have marked those 800 invoices, which is real human hours. Unsupervised learning needs none of that, which is why it gets used for exploring data nobody has had time to categorize — and why its answers need interpreting rather than trusting.
The unsupervised side gets a part of its own in this series, covering what clustering actually does and what inference means once a model is trained — that is Day 2, linked at the foot of this page. And if you are heading for the engineer-level exam, supervised learning in depth — classification and regression is the one to read after that.
Training and inference: the distinction that shows up on your bill
Two words that sound like jargon and are actually the most practical thing on this page.
Training is building the model from data. It happens occasionally, it is computationally enormous, and it is the part with the dramatic numbers attached.
Inference is using the finished model to answer one question. It happens constantly, and each individual request is cheap.
Here is why that matters the moment you touch a cloud platform: they are billed differently, and the one that surprises people is inference. Training is a known, chunky, deliberate cost you plan for. Inference is a small cost multiplied by however many times your application asks a question — and applications ask far more often than anybody estimates. When somebody says their AI costs are running away, they have almost never trained too much. They are doing more inference than they thought.
The exam leans on the distinction because it decides which service and which pricing model fits a scenario. Jake's version, which is not wrong: "Training is buying the van. Inference is the fuel. Everyone budgets for the van."
Which AWS exam — and the September 2026 dates
This is the part that changes what you should do this week, so it comes with numbers rather than encouragement.
| AI Practitioner | ML Engineer – Associate | |
|---|---|---|
| Level | Foundational | Associate |
| Questions | 65 | 65 |
| Time | 90 minutes | 130 minutes |
| Cost | $100 USD | $150 USD |
| Expected experience | None formally required | About a year with SageMaker and ML engineering |
| Who it is for | Understanding and discussing AI — analysts, managers, sales, developers new to it | Building it — developers, data engineers, MLOps, data scientists |
| Valid for | 3 years | 3 years |
They are not steps on a ladder. They are exams for two different jobs, and plenty of people only ever want the first.
⚠️ If you are booking the ML Engineer exam, read this first
The exam is changing version, and the window is short:
• 1 September 2026 — registration opens for the updated version (English only at first)
• 28 September 2026 — last day to take the current version in English
• 29 September 2026 — beta delivery of the updated version begins: 85 questions, 170 minutes, $75 USD
So: if you have already been studying, sit it before the cutoff and your preparation matches your exam. If you are starting from scratch now, aim at the new version instead — there is no sense learning to a syllabus that is about to be replaced. And check what any course you buy is actually preparing you for, which is the mistake this post opened with.
What the updated version adds
The changes tell you where the industry has moved, which makes them worth reading even if you never sit the exam:
- Building and deploying generative AI solutions — not just traditional models.
- Fine-tuning foundation models — adapting a general model to your own domain.
- Retrieval-augmented generation — the standard pattern for answering from your own documents.
- Agents and workflow orchestration — the sixth circle, now examinable.
- Expanded Amazon Bedrock coverage, and foundation model selection and operationalization.
- Responsible AI across both traditional and generative workloads.
No new domains were added; the existing ones were brought up to date. Two years ago this exam was largely about training your own models. It now assumes you will more often be adapting somebody else's — which is exactly the shift described in circle four.
And the exam that no longer exists
The older Machine Learning – Specialty certification has been retired, and its ground was divided between these two: the conceptual and governance side went to the AI Practitioner, and the practical build-and-deploy side to the ML Engineer Associate.
This matters commercially, not just administratively. A course still advertising preparation for the specialty exam is a strong signal that nobody has maintained it — which is the cheapest quality check available to you before spending money, and it takes ten seconds.
Where the AI Practitioner exam actually spends its questions
| Domain | Weight | Which circle |
|---|---|---|
| Fundamentals of AI and ML | 20% | Circles 1–3, plus the three ways of learning |
| Fundamentals of generative AI | 24% | Circle 5 |
| Applications of foundation models | 28% | Circle 4 — the largest domain |
| Guidelines for responsible AI | 14% | Bias, fairness, transparency, harm |
| Security, compliance and governance | 14% | Data protection, access, auditability |
Two things fall straight out of that table.
More than half the exam is generative AI and foundation models — 52 percent between them. If you have been treating those as the fashionable part and the "real" machine learning as the substance, that is the wrong way round for this exam.
Responsible AI and governance are 28 percent combined, the same as the largest technical domain. This is the part people skim, because it has no commands in it. It is more than a quarter of your score.
🔬 How this was verified
Every exam number and every date on this page was read from AWS's own certification pages and the AWS training and certification blog on 9 August 2026 — question counts, durations, prices, domain weightings, and the September schedule for the updated engineer exam. These figures move. Check the official page before you pay for anything, and if something has shifted since we wrote this, tell us and we will fix it.
One deliberate omission: we have not given a retirement date for the old Machine Learning – Specialty exam. That it has been retired and split into these two is certain; the exact day we saw only second-hand, so it is not stated as fact here. Exam details do change — check the official page before you pay, and if you find something on this page that has moved, tell us and we will correct it.
The words you will meet in week one
Not a full glossary — the ones that appear immediately and are usually left undefined.
| Term | What it means, plainly |
|---|---|
| Model | The thing produced by training. A file of numbers that turns an input into an output. |
| Parameters | Those numbers. "Billions of parameters" means the file is very large, not that it is wise. |
| Token | A chunk of text, roughly a short word. Usage is billed in these, so they appear on invoices. |
| Prompt | The input you give a generative model, including any instructions and context. |
| Hallucination | A confident, fluent, wrong answer. Not a malfunction — a consequence of how these models work. |
| Fine-tuning | Further training of an existing model on your own data to specialize it. |
| RAG | Look the facts up first, hand them to the model with the question. Cheaper than fine-tuning and easier to keep current. |
| Bias | Systematic unfairness learned from the data. A responsible-AI topic, and examinable. |
Hallucination is the one to internalize on day one. A generative model produces text that is plausible, and plausible and true are different targets. It is not broken when it invents a citation; it is doing the thing it does. That single realization is worth more than any prompt trick, and it is the reason retrieval-augmented generation exists at all — if you want an answer grounded in real documents, you have to hand it the documents.
🙋♂️ Jake's Reality Check
"Right, two honest questions. Is a hundred-dollar certificate going to get me anything? And is this stuff coming for my job?"
On the certificate: it will not get you a job on its own, and anybody promising that is selling something. What it does is give you a syllabus written by people who know the material, a deadline that makes you finish, and vocabulary you can use with a customer without hedging. For a hundred dollars and a few weeks of evenings, that is a fair trade — and for you specifically, the value is walking out able to answer the questions people are already asking you across the counter.
On the job: not the one where somebody carries a broken laptop into a shop and needs a person to work out what is wrong with it and be trusted with their photographs. The part of your work that is answering the same setup question forty times a year is more exposed, and that is worth knowing rather than being reassured about. The people doing well out of this are mostly not the ones who feared it or dismissed it — they are the ones who learned enough to tell which of the two their own work is.
What Day 1 should leave you able to say
Close the page and answer these out loud. If one sticks, that is the paragraph to reread — it takes two minutes and it is worth more than rereading everything.
- Why is all machine learning AI, but not all AI machine learning?
- What does the "deep" in deep learning refer to?
- What makes a foundation model different from a model trained for one task?
- Is generative AI a technology or a use? Why does the distinction matter?
- Which of the three learning types needs somebody to label the data first?
- Which costs more over a year on a busy application — training or inference?
- Which of the two AWS exams fits the work you actually do?
That last one is the only question on the list with money attached, which is why it has a table of its own further up.
Where this series goes next
Day 2 goes into the unsupervised side — clustering and inference, and how a machine knows that A means A. After that the track splits by which exam you are heading for. For the engineer route, data fundamentals, ingestion and transformation comes first, because on real projects the data work is most of the work, and problem formulation — does your business even need machine learning is the one that saves people from building something they did not need.
If you are new to AWS itself rather than to AI, start instead with the free AWS learning series, because the AI material assumes you know what a service and a region are.
Questions people actually ask about this
What is the difference between AI and machine learning?
Machine learning is one part of artificial intelligence, not a synonym for it. Artificial intelligence is the whole field of getting machines to do things that normally need human thinking. Machine learning is the specific approach where you show a system many examples and let it work out the pattern, instead of writing the rules yourself. All machine learning is AI. Plenty of AI, such as older rule-based systems, is not machine learning.
What is a foundation model in simple terms?
It is a very large model trained on a very broad range of data, so that it can be adapted to many different tasks rather than built for one. Older models were trained for a single job, such as spotting spam. A foundation model is trained once at enormous expense and then pointed at summarizing, translating, answering or writing without being retrained from scratch. That reusability is the whole reason the term exists.
Is generative AI the same as machine learning?
No, it is a use of it. Generative AI means using a model to produce new content, such as text, images or code, rather than to produce a label or a number. The underlying machinery is deep learning and, these days, foundation models. So generative AI sits inside machine learning, which sits inside artificial intelligence. Getting that nesting right makes most of the vocabulary fall into place.
Which AWS AI certification should I start with?
Start with the AI Practitioner if you want to understand and talk about AI on AWS without building it, which suits analysts, managers, salespeople and developers new to the area. Choose the Machine Learning Engineer Associate if you will actually build, train and deploy models, and expect it to want real hands-on experience. They are different levels for different jobs rather than steps you must climb in order.
How many questions is the AWS AI Practitioner exam and what does it cost?
It is 65 questions in 90 minutes and costs 100 US dollars, and it is a foundational-level exam with no formal prerequisites. The certification is valid for three years. It can be taken at a test center or online with a proctor. If you are completely new to AWS itself, working through the cloud fundamentals material first makes the AI content much easier to absorb.
What are the AWS AI Practitioner exam domains?
There are five. Fundamentals of AI and machine learning at 20 percent, fundamentals of generative AI at 24 percent, applications of foundation models at 28 percent, guidelines for responsible AI at 14 percent, and security, compliance and governance for AI solutions at 14 percent. More than half the exam sits in the two generative AI and foundation model domains, which tells you where to spend your study time.
Is the AWS Machine Learning Engineer exam changing?
Yes, and the dates matter if you are booking a seat. Registration for the updated version opens on 1 September 2026, the last day to take the current version in English is 28 September 2026, and beta delivery of the new version begins on 29 September 2026. If you have been studying for the current version, sitting it before that cutoff means your preparation matches the exam you take.
What is new in the updated ML Engineer Associate exam?
The update brings the exam in line with how machine learning work is actually done now. It adds building and deploying generative AI solutions, fine-tuning foundation models, retrieval-augmented generation architectures, orchestrating AI agents and complex workflows, and expanded coverage of Amazon Bedrock, along with responsible AI practices across both traditional and generative workloads. No new domains were added, the existing ones were updated.
Should I take the beta version of an AWS exam?
It is cheaper and it is slower to give you a result. The beta sitting of the updated ML Engineer exam is 85 questions in 170 minutes for 75 US dollars, against 65 questions in 130 minutes for 150 dollars on the current version. You wait longer for your score because the results are held until the scoring is finalized, and study material for a brand new version is thin. Worth it if you are confident and patient.
What happened to the AWS Machine Learning Specialty certification?
It has been retired, and its ground was split into two. The AI Practitioner covers the foundational and conceptual side for people who need to understand and discuss AI, and the Machine Learning Engineer Associate covers the practical build-and-deploy side. If you find a course still selling preparation for the specialty exam, that is a strong signal the material has not been maintained.
What is the difference between training and inference?
Training is the process of building the model from data, which is done occasionally and is computationally expensive. Inference is using the finished model to answer a question, which happens constantly and is cheap per request. The distinction matters on any cloud platform because the two are billed differently, and a system whose costs are running away is usually doing far more inference than anyone estimated rather than too much training.
What is RAG and why does it keep coming up?
Retrieval-augmented generation means looking up relevant information first and handing it to the model along with the question, instead of relying on what the model absorbed during training. It is the standard way to make a general model answer questions about your own documents without retraining it. It appears constantly because it is cheaper than fine-tuning, easier to keep current, and it lets you show where an answer came from.
Do I need to know how to code for the AI Practitioner exam?
No. It is aimed at people who are familiar with AI and machine learning solutions without necessarily building them, which explicitly includes business analysts, product managers and salespeople. You need the concepts, the vocabulary, an understanding of which AWS service fits which job, and the responsible AI and governance material. Coding matters for the engineer-level exam, not this one.
Is an AI certification worth it, honestly?
It is worth it as a structure to learn against and as something recognizable on a profile, and it is not a job on its own. The foundational exam is a reasonable investment for a hundred dollars if you need the vocabulary for your current work. The engineer-level one carries more weight because it expects experience you cannot fake. What neither will do is substitute for having built something you can talk about.
How long does it take to prepare for the AI Practitioner exam?
For somebody already working in tech but new to AI, a few weeks of steady evening study is a realistic range, and less if you already use these tools daily. It is a foundational exam and it is not trying to catch you out. The material that takes longest is usually the responsible AI and governance content, because it is the part people skim and it carries a similar weight to security and compliance.
What Jake did with the four hundred dollars
He closed the tab, booked the hundred-dollar foundational exam for eight weeks out, and put the difference towards a second monitor for the shop, which he had wanted for two years.
"And if I decide I want the engineer one later?"
"Then you'll want the new version, and by the time you get there the study material will exist. Right now you'd be learning to a syllabus that expires in September."
"Feels like the actual skill here is working out what's current."
"In this field, quite often, yes."
Which is a reasonable note to end Day 1 on. The five circles will still be true in ten years — the nesting does not change. The exam codes, the prices and the cutoff dates will have moved twice by then. Learn the first properly, and check the second before you pay for anything.
- Day 2: clustering, inference, and how a machine knows that A means A
The unsupervised half of the three learning types, worked through properly. - Supervised learning in depth: classification and regression
The next step if you are heading for the engineer-level exam rather than the foundational one. - Problem formulation: does your business even need machine learning?
The question worth asking before any of the technical work starts. - What AWS is, and how people actually make money with it
Background if the cloud part, rather than the AI part, is what is new to you.
Revision note. Originally published 16 October 2024 as Day 1 of this series, and rewritten on 9 August 2026. The original was written while the AI Practitioner exam was still new and the generative AI landscape was changing month to month, which is a hard moment to write a foundations post into. Two years on, enough has settled to say it properly: the exam exists with published domains and weightings, the old Machine Learning – Specialty has been retired and split in two, and the engineer-level exam is about to move to a new version that adds fine-tuning, retrieval-augmented generation and agents — the things that were research talk when the first version of this post went out. So this rewrite keeps the beginner promise and adds the two things the original could not have: a concept ladder that makes the vocabulary stick, and the current exam numbers and dates, checked against AWS on the day of writing. If you are studying for either of these and something here does not match what you see when you book, please write in through the contact page — certification details move, and a reader who spots it first saves everyone who arrives after them.