Generative AI Explained: From Discriminative AI to Agentic AI

Introduction

Artificial Intelligence (AI) has changed a lot over time. In the beginning, AI was mainly used to follow rules and make predictions. Later, Machine Learning and Deep Learning helped computers learn from data. Today, Generative AI can create new content like text, images, audio, videos, and code.

This blog explains the journey from Discriminative AI to Generative AI and Agentic AI.

What is Discriminative AI?

Before Generative AI, we mainly used Discriminative AI models.

Discriminative AI focuses on finding the difference between classes of data. It learns patterns from existing data and predicts the correct result.

The main idea is:

P(Class | Data)

It means:

"What is the probability of a class when we have given data?"

Examples:

• Is this email spam or not?
• Is this image a cat or a dog?
• Is a transaction fraud or normal?

Limitations of Discriminative AI

Discriminative AI is good at classification and prediction, but it cannot create new content.

It can answer:

"Is this a cat?"

But it cannot create:

"Generate a new image of a cat."

To solve this limitation, Generative AI was introduced.

What is Generative AI?

Generative AI is a type of AI that can create new content based on the data it has learned from.

It can dynamically generate:

• Text (articles, code, answers)
• Images (art, product photos, designs)
• Audio (voice, music)
• Video (animations, synthetic video)

This was a big shift. Instead of just predicting a label, the AI now had to understand and produce content.

The Four Core Architectures of Generative AI

Generative AI is not built on a single model. There are four main architectures, and each one is good at different things.

1. GANs (Generative Adversarial Networks)

GANs are mainly used for image generation.

They have two parts:

• Generator: Creates new content
• Discriminator: Checks if content looks real

The generator improves by trying to fool the discriminator.

Example:

• Creating realistic human faces
• Image generation

Known models: StyleGAN, BigGAN

Note: Earlier versions of DALL-E were associated with GAN research, but modern DALL-E (2 and 3) is built on Transformer + Diffusion architecture.

2. VAEs — Variational Autoencoders

VAEs work by compressing data into a small representation and then reconstructing it. They are good at learning the structure of data.

Best for: Data compression, reconstruction, and generating variations of existing data

Known models: Beta-VAE, VQ-VAE

3. Transformers

This is the architecture that changed everything. Transformers use a mechanism called attention — they figure out which parts of the input matter most when producing an output.

Best for: Text, audio, video, and image generation. Known LLMs built on this:

• GPT series (OpenAI)
• Claude (Anthropic)
• Gemini (Google)
• LLaMA (Meta)
• DALL-E 3 (OpenAI) — multimodal Transformer

4. Diffusion Models

Diffusion models work by learning to remove noise from an image step by step. They start with pure noise and slowly clean it into a final image.

Best for: High-quality, detailed image generation

Known models: Stable Diffusion, Midjourney, DALL-E 3

Today, Diffusion Models have largely replaced GANs for image generation tasks because they produce more stable and higher-quality results.

How Did AI Evolve Over Time?

Stage                                     : What it did

Early / Rule-Based AI     = Followed fixed, hand-written rules

Machine Learning (ML)     =    Learned patterns from data to make predictions

Deep Learning (DL)    =    Understood complex patterns using neural networks

Generative Models     =    Created new content (GANs, VAEs, Diffusion)

Foundation Models    =     Large pre-trained models used for many tasks (GPT, Claude, Gemini) 

Popular LLMs and What They Are Best At

There are many LLMs available today. Each one is built for a specific use case.

Model	Best For
ChatGPT	Dynamic conversations and general-purpose responses
Gemini	Research-heavy tasks and Google ecosystem integration
Claude	Long-context tasks, safe and natural long-form writing
Jasper	Marketing content and brand copywriting
Rytr	Blog posts, emails, and SEO content
Resoomer	Text summarization
DeepL / Google Translate	High-quality language translation
NOLEJ	Education-focused AI content
Watson AIOps	Analyzing system logs, data streams, and detecting anomalies
Midjourney	Artistic, high-detail, and visually aesthetic image generation

Generative AI vs Agentic AI

Generative AI and Agentic AI are related but different.

Generative AI	Agentic AI
Creates content	Takes actions
Generates text, images, and code	Makes decisions and completes tasks
Reactive	More proactive
Example: Writing an email	Example: Booking a flight

How They Work Together

Agentic AI does not replace Generative AI — it uses it.

Think of it this way:

• Generative AI is the brain — it understands the goal and creates a plan
• Agentic AI is the hands — it executes that plan step by step in the real world