Lesson 1.1: Introduction to Generative AI

You’re a project manager at a healthcare startup. Your team built an AI tool to help doctors diagnose rare diseases. Last month, it made a dangerous mistake: the AI misread a patient’s symptoms as “seasonal allergies” instead of a life-threatening heart condition. The model relied on outdated medical guidelines and couldn’t access new research.

This failure illustrates a critical challenge with traditional AI systems: they often operate with static knowledge, like a textbook that never gets updated. By the end of this lesson, you will understand how modern generative AI works, why it is transforming industries, and how to avoid disasters like this by grounding AI in dynamic, real-world data.

Generative AI is the creative powerhouse, discriminative AI is the expert classifier

Generative AI creates new content — text, images, code, or music — by learning patterns from vast datasets. Think of it as an artist or inventor.

Examples include:

• ChatGPT, which writes essays, poems, or software code.
• DALL-E, which generates images from prompts like “a penguin wearing a top hat.”
• GitHub Copilot, which auto-completes code and boosts developer productivity by 55%.

Discriminative AI analyzes existing data to label, classify, or predict outcomes. It acts more like a detective or judge.

Examples include:

• Spam filters like Gmail’s junk mail detection.
• Facial recognition systems that unlock your phone.
• Credit scoring models predicting loan defaults.

The analogy is simple:

• Generative AI is like a chef inventing a new recipe.
• Discriminative AI is like a food critic rating dishes.

You will use generative AI when you want to create something novel. Use discriminative AI when you want to detect or classify.

The transformer revolution changed AI forever

Before 2017, AI systems mostly relied on sequential models like Recurrent Neural Networks (RNNs) and Long Short-Term Memory networks (LSTMs). These processed data word-by-word, like reading a sentence left to right. They struggled with long sentences and complex dependencies—for example, understanding that in “The cat the dog chased was brown,” the cat is brown.

Earlier AI systems also used handcrafted rules. Chatbots could only answer questions they were explicitly programmed for. Ask “What’s the meaning of life?” and they would fail.

Then came the 2017 paper “Attention Is All You Need” by Vaswani et al., introducing Transformers. This architecture changed AI fundamentally by using:

• Self-attention, which lets models weigh the importance of every word in a sentence simultaneously. For example, in “She gave him the book because it was insightful,” the model understands that “it” refers to “book.”
• Parallel processing, enabling models to look at entire sentences at once instead of one word at a time. This allows faster training on massive datasets like Wikipedia, books, and code repositories.

The result: models like GPT-3 (2020) can generate human-like text, solve math problems, and write code.

To visualize:

• Pre-transformers are like reading a book with a magnifying glass—one word at a time.
• Transformers let you view the entire page at once.

This breakthrough underpins modern generative AI’s creativity and fluency.

Why generative AI matters economically

McKinsey estimates that generative AI could add $4.4 trillion annually to the global economy by:

• Automating repetitive tasks like drafting emails and generating reports.
• Accelerating software development — GitHub Copilot, for instance, writes 40% of code in some projects.
• Enhancing creativity in advertising, product design, and more.

GitHub Copilot’s case is instructive:

• Developers spent hours writing boilerplate code.
• Copilot suggests code snippets in real-time.
• 55% of developers reported faster task completion.
• 75% felt more focused on creative problem-solving.

Generative AI is not about replacing humans but augmenting human potential. For example:

• A marketer drafts campaign ideas with ChatGPT, then refines them.
• A radiologist uses AI to flag anomalies in X-rays before making the final diagnosis.

Fixing AI disasters with dynamic knowledge: Retrieval-Augmented Generation (RAG)

What went wrong with the hospital AI? It used static training data — outdated guidelines — and had no way to access new research.

The solution is Retrieval-Augmented Generation (RAG), which combines generative AI with real-time data retrieval:

1. Retrieval: The system queries updated medical databases like PubMed or FDA alerts. Think of it as giving the AI a library card instead of a single textbook.
2. Generation: The AI drafts a diagnosis using the latest information.
3. Safety guardrails: Rules ensure critical symptoms trigger escalation, for example, “If symptoms include chest pain, escalate to a cardiologist.”

The impact is real:

• The Mayo Clinic reduced diagnostic errors by 40% using a similar approach.
• Google’s Med-PaLM 2 answers medical questions with 85% accuracy by grounding responses in clinical guidelines.

This pattern is essential. No matter how powerful the model, if it cannot access fresh, authoritative data, it risks dangerous mistakes.

Test your understanding

1. Generative AI is best used for:

   a) Detecting credit card fraud

   b) Writing a blog post

2. The self-attention mechanism was introduced in:

   a) 2010

   b) 2017

3. GitHub Copilot improved developer productivity by:

   a) 25%

   b) 55%

Common pitfalls and how to avoid them

Pitfall 1: Blind trust in AI outputs

Example: A lawyer used ChatGPT to draft a court filing—it cited six fake court cases.

Fix: Treat AI as a first draft tool, not a final authority. Always fact-check critical outputs.

Pitfall 2: Ignoring data updates

Example: A chatbot trained in 2021 told users COVID-19 vaccines were “experimental” in 2023.

Fix: Use Retrieval-Augmented Generation to pull live data.

Pitfall 3: Overlooking bias

Example: A hiring tool penalized resumes with the word “women’s,” e.g., “women’s chess club captain.”

Fix: Audit training data for biases using tools like IBM AI Fairness 360.

Homework: Hands-on exploration

Task 1: Experiment with generative AI

• Use ChatGPT, Gemini, and Claude to draft a 200-word product description for “solar-powered headphones.” Compare the outputs.
• Use DALL-E (integrated within ChatGPT), MidJourney, or MetaAI (integrated within WhatsApp or Instagram) to generate an image of “a robot gardening on Mars.”

Task 2: Compare with discriminative AI

• Use Google’s Vision API to identify objects in a photo of your workspace.
• Use Hugging Face’s sentiment analyzer to check if your ChatGPT product description sounds positive.

Reflect on:

• How accurate were the AI-generated outputs?
• Where could these tools go wrong in a real business setting?

Where to go next

• Understand tokenization and context windows: Tokens, Context Windows, and Scaling Laws
• Explore fine-tuning and RAG approaches: Fine-Tuning vs. Retrieval-Augmented Generation (RAG)
• Learn about compliance and ethics in AI: Compliance in Healthcare AI Systems

PL alumni now work at Razorpay, Swiggy, Flipkart, PhonePe, Microsoft, and other leading Indian tech companies.