What is Generative AI?

Introduction

Generative AI is the family of AI models that create brand-new content — text, images, audio, video, code — instead of only classifying or predicting. ChatGPT generating an essay, Midjourney painting a picture, GitHub Copilot writing code, ElevenLabs cloning a voice — all are Generative AI.

Beginner analogy: Classic ML is a judge ("is this email spam?"). Generative AI is an artist ("write me a haiku about the sea"). Both learn from data, but generative models produce data instead of labelling it.

Understanding the topic

Core concepts to understand:

• Generates new artefacts: text (LLMs), images (diffusion), audio (TTS), video, code.
• Trained on huge corpora — the entire public internet, code, books, images.
• Driven mostly by transformer architectures (since 2017's Attention Is All You Need).
• Output is probabilistic — the same prompt rarely produces identical text twice.
• Capable of zero-shot generalisation: solving tasks it was never explicitly trained on.

Syntax reference

Visual workflow / architecture:

Training data (text, code, images)
       │
       ▼
┌──────────────────────────┐
│   Transformer Model      │
│  (billions of params)    │
└─────────────┬────────────┘
              │ learns patterns
              ▼
┌──────────────────────────┐
│   Generative Model       │
│  GPT · Claude · Gemini   │
└─────────────┬────────────┘
              │ given a prompt
              ▼
       NEW content
   (text · image · code)

Real-world use

OpenAI's GPT-4, Anthropic's Claude, Google Gemini, Meta Llama, Stable Diffusion and Sora (video) are all Generative AI. Stripe uses GPT to triage support tickets; Duolingo uses GPT-4 to roleplay language conversations; Microsoft embeds Copilot in every Office app.

Best practices

• Always set temperature appropriately: 0 for deterministic tasks, 0.7+ for creative writing.
• Validate generated output — it can sound confident and still be wrong (hallucination).
• Cache common prompts — generation is the slow, expensive step.

Common mistakes

• Treating generative output as factual — always verify in regulated domains.
• Forgetting cost: every token in & out is billed.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Define Generative AI.
• Q2. How is it different from a classifier like spam detection?
• Q3. Name three production-grade generative models.
• Q4. What is a hallucination and why does it happen?