AI vs LLMs
Hierarchy
Definitions
- AI → broad field where machines do "smart" tasks
- Discriminative AI → Part of AI that classifies or predicts (labels data)
- Generative AI → Part of AI that creates new content (text, images, etc.)
- LLM → Part of AI that is trained on massive text data to generate text
- Image Models → Part of AI that is trained on massive image data to generate images
Examples
| Term | Real-world use | How it works |
|---|---|---|
| Discriminative AI | Spam filters in Gmail, Fraud detection, image classification | Map input to a label |
| Generative AI | ChatGPT, DALL·E | Learn patterns to create new data |
| LLM | GPT-4/5, Anthropic, LLaMA | Predicts next token to create language |
| Image models | Stable Diffusion, Midjourney | Learns features to create new pixel arrangaments |
Generative vs Discriminative AI
| Aspect | Discriminative AI | Generative AI |
|---|---|---|
| Goal | Classifies / labels | Creates |
| Output | Category, prediction | Text, image, audio, code |
| Examples | Logistic regression, object detection | LLMs, GANs, diffusion models |
Agentic AI
We can also define another type of AI → Agentic AI
Agentic AI performs multi-step tasks on your behalf by orchestrating tools, APIs, and workflows.
| Agent | What you ask | What it does |
|---|---|---|
| 🤖 GitHub Copilot | "Implement this feature" | Creates a branch → generates code → runs checks → pushes the commit |
| ✈️ Travel assistant | "Book me a trip to Berlin" | Searches flights → compares prices → fills booking form → confirms reservation |
| 📅 Calendar assistant | "Schedule a team sync" | Checks everyone's availability → proposes a slot → creates the event → sends invites |
| 🚀 DevOps deployer | "Ship the new release" | Runs tests → builds artifacts → deploys to staging → monitors health → rolls back on failure |
| 📊 Data analyst | "Summarise last month's sales" | Queries the database → runs analysis → generates charts → files a report |
| 🎧 Support agent | "Handle incoming tickets" | Triages issues → drafts replies → escalates complex cases → creates follow-up tasks |