What Is AI, and What Is Generative AI?

What You Will Learn

This lesson covers three things. By the end you should be able to define AI accurately in plain language, distinguish traditional software from machine learning and generative AI, and calibrate your own expectations. Both over-hype and under-estimation are equally unhelpful when it comes to building real systems.

Two Ways to Build Intelligence

Every piece of software that makes decisions is built using one of two fundamentally different approaches. Understanding this distinction is the foundation of everything that follows in this course.

Traditional Software: Rules Written by Hand

In traditional software, a human developer writes every decision in advance. If the user clicks X, do Y. Always. A spam filter built this way works by checking emails against a handcrafted list of suspicious keywords. This approach is predictable and reliable — but rigid. If a new type of spam appears that does not match any handcrafted rules, the filter fails.

AI / Machine Learning: Rules Learned from Examples

Machine learning takes a completely different approach. Instead of writing rules by hand, you show the system millions of labelled examples, and it extracts the patterns itself. A spam filter built this way was trained on millions of emails labelled as spam or not spam. It learned the patterns from the data. This makes the system flexible and adaptive — it improves with more data and can handle situations the developer never anticipated.

What Exactly Is a Model?

A model is not a program in the traditional sense. It is a very large set of mathematical patterns — specifically, numerical weights extracted from data during training. You feed the system millions of examples. The system adjusts its internal parameters — billions of them — until it can reliably predict the correct output for each input. Once training is complete, those adjusted parameters are saved. That saved set of parameters is the model.

From Machine Learning to Generative AI

Machine learning has been around for decades. The answer to what changed recently lies in a specific sequence of breakthroughs. In 2012, AlexNet demonstrated that deep learning could outperform every hand-engineered system. In 2017, the Transformer architecture was introduced in 'Attention Is All You Need' — every modern LLM is a transformer. From 2022 onwards, these models scaled to billions of parameters and were exposed to vast portions of the public internet during training.

Traditional machine learning classifies and predicts — is this email spam? Generative AI creates new content — text, images, code — that did not exist before. The output is generated token by token based on patterns learned during training.

What LLMs Are Good and Bad At

Why Healthcare Is One of the Best Domains for LLMs

Healthcare is unusually text-heavy. A single hospital generates millions of words of structured and unstructured text every day — clinical notes, discharge letters, referral letters, insurance requests, adverse event reports, and research literature. Almost all of it requires reading, understanding, and extracting specific information.

This is precisely what LLMs are good at. A model that reads a 12-page clinical note and extracts the three most important follow-up actions is genuinely useful. A model that translates a dense discharge letter into plain language a patient can actually follow is solving a real, documented problem — failure to follow discharge instructions is one of the leading causes of 30-day hospital readmission.

The constraint in healthcare is not whether LLMs can help — it is whether they can be deployed safely, with appropriate human oversight and privacy protection. You will encounter both the opportunity and the constraint throughout this course.