Topics
What is Generative AI?
My notes on understanding what GenAI actually is — and how it’s different from the AI we’ve always heard about.
🤖 Traditional AI vs Generative AI vs Agentic AI
AI has evolved in distinct waves — each one expanding what machines can do.
| Traditional AI | Generative AI | Agentic AI | |
|---|---|---|---|
| What it does | Classifies, predicts, decides | Creates new content | Takes actions autonomously to achieve goals |
| Output | A label, a number, a decision | Text, images, audio, video, code | Completed tasks, executed workflows |
| Example task | “Is this email spam?” | “Write me an email” | “Check my inbox, summarize unread emails, and draft replies” |
| How it learns | From labeled examples | From massive amounts of raw data | Combines GenAI with tools, memory, and planning |
| Mental model | A judge | An artist | An employee |
🧠 Traditional AI — The Old Guard
Traditional AI (also called discriminative AI) was designed to draw boundaries and make decisions.
Think of it as teaching a system to sort things into buckets:
- Is this tumor malignant or benign?
- Will this customer churn?
- Which face is in this photo?
It’s powerful — but it can only work within what it’s been explicitly trained to classify. It has no ability to imagine or generate anything new.
✨ Generative AI — The Creative Shift
Generative AI flips the script. Instead of asking “what is this?”, it asks “what comes next?”
At its core, a generative model learns the patterns and structure of data so deeply that it can produce brand new examples that look like they came from the same source.
- Feed it millions of books → it learns to write
- Feed it millions of images → it learns to draw
- Feed it millions of lines of code → it learns to program
The key insight: generation is a form of deep understanding. To create something convincing, the model must have truly learned the underlying structure of the domain.
📦 What Can Generative AI Generate?
| Modality | Examples |
|---|---|
| Text | Essays, summaries, code, conversations |
| Images | Art, photos, diagrams |
| Audio | Music, voice, sound effects |
| Video | Clips, animations |
| Structured data | Tables, synthetic datasets |
| Multi-modal | Image + text together (e.g., describe an image, generate from a description) |
🕹️ Agentic AI — The Autonomous Leap
Agentic AI is the next frontier. It takes Generative AI and gives it autonomy, tools, and a goal.
Instead of just responding to a prompt, an AI agent can:
- Plan — break a big goal into smaller steps
- Act — use tools like web search, code execution, or APIs
- Remember — retain context across steps
- Self-correct — evaluate its own output and try again if something fails
Think of it this way:
- Generative AI answers “How do I book a flight?”
- Agentic AI actually books the flight for you
The Building Blocks of an Agent
| Component | What It Does |
|---|---|
| LLM (the brain) | Reasons, plans, and generates responses |
| Tools | Lets the agent interact with the world (search, code, APIs) |
| Memory | Stores context — short term (within a task) and long term (across sessions) |
| Feedback loop | Agent checks its own output and iterates |
Real World Examples
- Coding agents — given a bug report, they read the code, identify the issue, write a fix, and run tests
- Research agents — given a topic, they search the web, read sources, synthesize findings, and write a report
- RAG pipelines — a form of agentic behavior where the model retrieves relevant knowledge before answering (this is what the DeepLearning.AI course covers!)
🔑 Key Concepts to Know
Foundation Models
Large models trained on vast, general datasets that can be adapted to many tasks. GPT-4, Claude, and Gemini are examples. They are the backbone of most modern GenAI applications.
Prompting
The way you communicate with a generative model — your input shapes its output. Unlike traditional AI where you feed it data, here you converse with it.
Emergent Behavior
As models get larger, they start exhibiting capabilities nobody explicitly trained them for — like reasoning, translation, or coding. This surprised even the researchers building them.
Hallucination
A known weakness — generative models can produce confident-sounding but completely wrong information. Understanding this is critical before building anything with GenAI.
💡 My Takeaways
- The shift from discriminative to generative is the fundamental mental model change
- GenAI doesn’t “know” things the way humans do — it predicts plausible outputs based on patterns
- The same model can write poetry, debug code, and explain quantum physics — that flexibility is what makes it so powerful and so different from anything before
- Hallucination is not a bug being fixed — it’s an inherent property to understand and design around
- Agentic AI is where things get really interesting — it’s the bridge between AI as a tool and AI as a collaborator
- RAG is a great first step into agentic thinking — the model retrieves before it responds, which is a simple but powerful form of tool use
❓ Questions I Still Have
- Where exactly is the line between a “large” and “small” model?
- How do foundation models get adapted for specific tasks? (fine-tuning — to explore next)
- How does RAG help reduce hallucinations?