AI, machine learning and deep learning

On the first Tuesday of October 2026, Noor Siddiqui set up a laptop in the small boardroom above the Calgary emergency hospital and wrote one question on the whiteboard: "What would we actually let a computer decide?" Cascade Ridge Veterinary Partners had hired her for nine months to assess where artificial intelligence could help its 11 clinics and its 24-hour hospital. The board had read about AI in every newspaper. Dr. Simone Kaur, the medical director, had read about it too, and she arrived with a firm opinion: "Nothing that touches a diagnosis without a vet in the loop."

Jonah Petrakis, the client services and scheduling analyst, was the other person at the table. He builds the booking reports that tell clinic managers how many appointments were kept, cancelled or missed, and he plans to sit the AWS Certified AI Practitioner exam (AIF-C01) in April 2027. Noor agreed to teach a weekly "AI hour" that would serve both purposes: Jonah would learn the vocabulary the exam expects, and the company would get a clear-eyed list of AI uses worth trying. This course follows those sessions. It covers the first domain of the exam guide, the fundamentals of AI and machine learning, and the AWS services that apply them.

What artificial intelligence means

Artificial intelligence (AI) is the broad field of building computer systems that perform tasks we associate with human intelligence: recognizing speech, understanding text, spotting patterns in images, making predictions, planning steps toward a goal. The term is old and wide. A chess program written with hand-coded rules in the 1980s counts as AI, and so does a modern system that writes a summary of a phone call.

What matters for practitioners is not the label but how a system gets its behaviour. There are two broad ways. In a rules-based system, people write the logic explicitly: if a vaccine was given on this date, the next dose is due 12 months later. In a learned system, people supply examples and an algorithm finds the pattern itself. Both are useful, and lesson 11 returns to the important question of which to choose.

Machine learning

Machine learning (ML) is the part of AI in which a system learns patterns from data instead of following rules a person wrote. Three words carry most of the meaning, and the exam uses them precisely.

• An algorithm is the learning procedure: the recipe that looks at data and adjusts numbers to reduce mistakes. Linear regression, decision trees and neural networks are families of algorithms.
• A model is the result of running an algorithm on data. It is the trained artifact, a file of learned parameters, that takes new input and produces an output such as a prediction or a category.
• Training is the process of fitting the model to historical data. Inference is using the trained model on new data to get an answer.

Noor put it in clinic terms. "Three years of appointment history is the data. The algorithm is the method we use to look for patterns in who misses appointments. The model is what comes out: a calculator that, given tomorrow's bookings, estimates which ones are likely to be missed. Training happens occasionally, on a schedule we choose. Inference happens every evening when the reminder list is built."

A model's output is almost always a prediction, not a fact. It is an estimate based on patterns in past data, and it carries uncertainty. A model that says an appointment has a 0.3 chance of being missed is not saying the client will not come; it is saying that among appointments that looked like this one in the past, about 3 in 10 were missed. Keeping that distinction in mind prevents a whole class of mistakes.

Deep learning and neural networks

A neural network is a family of ML models loosely inspired by the brain. It is built from layers of simple units, often called neurons, each of which multiplies its inputs by learned weights, adds them up and passes the result through a simple function. Stacking many layers lets the network represent complicated relationships. Deep learning is machine learning with neural networks that have many layers.

Deep learning is what made modern computer vision, speech recognition and language models possible. It shines on unstructured data such as images, audio and free text, where the useful features are hard for people to describe in rules. Its costs are real: deep models usually need large amounts of data and computing power to train, and they are harder to explain than a simple model. For a table of appointment records, a simpler model is often just as accurate and much easier to justify.

Areas of AI you will meet

Several application areas appear repeatedly in the exam and in practice.

• Computer vision lets systems interpret images and video: detecting objects, reading text in a photo, judging whether a radiograph is in focus.
• Natural language processing (NLP) works with human language in text: classifying a review as positive or negative, extracting names and dates, translating a letter.
• Speech recognition converts spoken audio to text, and speech synthesis turns text into natural-sounding speech.
• Recommendation systems suggest items a person is likely to want, and forecasting predicts future values such as next month's appointment volume.
• Fraud and anomaly detection flags transactions or events that do not fit normal patterns.

Generative AI, foundation models and LLMs

Generative AI is a branch of deep learning in which models produce new content, such as text, images, audio or code, instead of only labelling or scoring existing data. Most generative systems today are built on foundation models (FMs): very large models trained on broad data that can be adapted to many tasks. A large language model (LLM) is a foundation model specialized in text: it predicts the next piece of text given what came before, and that simple objective, at enormous scale, yields the ability to summarize, answer questions and draft documents. AWS-112 covers how these models work; for now, the important point is where they sit. Generative AI is a kind of deep learning, which is a kind of machine learning, which is one way of building AI.

Agentic AI

The newest term in the exam guide is agentic AI. An AI agent is a system that does not just answer one request but pursues a goal over several steps: it plans, calls tools such as a booking system or a search service, observes the result and decides what to do next. A chatbot that answers "what are your hours?" is generative AI. A system that reads a client's message, checks the schedule, proposes three appointment slots and books the one the client picks is agentic. Agents raise the stakes because they take actions, not just produce text, so permissions and human review matter even more. AWS-113 covers agents and tool use in depth.

Bias, fairness and fit: three words to watch

Three more terms from the exam guide appear in every later lesson, so Noor introduced them early. Bias has two meanings. In statistics, it is a model's systematic error, the tendency to be consistently wrong in one direction, which lesson 6 covers. In everyday use, and in responsible AI, it means outcomes that are systematically unfair to some group. Fairness is the goal of making a model's benefits and errors fall evenly across groups that should be treated alike. Fit describes how well a model's learned patterns match the data: a model can underfit, missing real patterns, or overfit, memorizing noise. All three are about the same underlying question: does the model's behaviour reflect reality, and reality for everyone it affects?

# Rules versus learning, in miniature.
# A rule: someone writes the logic.
def next_rabies_due(months_since_last):
    return "due" if months_since_last >= 12 else "not due"

# Learning: the "rule" (a threshold) is found from examples.
history = [(1, 0), (2, 0), (3, 0), (8, 1), (9, 1), (12, 1)]  # (days booked in advance, missed? 1=yes)

best_t, best_err = None, None
for t in range(1, 13):
    errors = sum((days >= t) != bool(missed) for days, missed in history)
    if best_err is None or errors < best_err:
        best_t, best_err = t, errors

print("rule says:", next_rabies_due(13))
print("learned threshold: booked", best_t, "or more days ahead -> higher miss risk; errors =", best_err)

rule says: due
learned threshold: booked 4 or more days ahead -> higher miss risk; errors = 0

The rule needed no data at all, and it is exactly right because vaccine schedules are set by protocol. The learned threshold came from examples, and it would change if the examples changed. That is the essential trade: learning can find patterns nobody wrote down, but it depends entirely on the data it was given.

A very short history

The ideas are older than most people assume. Researchers coined the term artificial intelligence in the 1950s, and the first neural network designs date from the same decade. For decades progress came in bursts followed by long "AI winters", when early promises outran what computers could do. Three things changed in the 2010s: far more digital data, much cheaper computing power (especially graphics processors, which happen to be good at the arithmetic neural networks need), and better training methods. Deep learning models began to beat older approaches at recognizing images and speech. The transformer architecture, published in 2017, made it practical to train very large language models, and the generative AI tools that reached the public from late 2022 are built on it. Knowing this history helps Jonah read news coverage calmly: the tools are new, but most of the underlying ideas, and many of the old warnings about overpromising, are not.

Where AWS fits

AWS offers AI at three levels, and the rest of this course keeps returning to them. At the top are managed AI services that solve a specific task through a simple API, such as turning speech into text or translating a document; no ML expertise is needed. In the middle are generative AI services such as Amazon Bedrock, which give access to foundation models from several providers through one API. At the base is Amazon SageMaker AI, a platform for building, training and deploying your own models, for teams with data scientists. Noor's advice was to start each idea at the highest level that does the job and move down only when a real need forces it: "Every step down gives you more control and more work."