The AI Landscape: From Narrow AI to Generative AI

Two Fundamentally Different Things AI Does

Not all AI works the same way or serves the same purpose. At the highest level, AI systems fall into two broad families: those that classify or predict, and those that create. Understanding this distinction helps you set the right expectations for any AI tool you encounter.

Discriminative AI takes an input and categorizes it or predicts a value. Examples: a spam filter that classifies emails as spam or not; a fraud detection system that flags suspicious transactions; a medical imaging tool that identifies potential tumors in a scan; a recommendation engine that predicts what you want to watch next. These systems are trained to recognize patterns in existing data and output a judgment or score.

Generative AI takes an input and creates new content: text, images, audio, video, code. ChatGPT generates text. Midjourney generates images. Suno generates music. These systems learned the patterns of their training data well enough to produce new examples that fit those patterns.

Both types have been transforming industries for years. The explosion of public attention since 2022 has been driven primarily by generative AI — because the outputs are immediately visible and shareable in a way that a fraud score or a classification label is not.

How Diffusion Models Create Images

Language models generate text by predicting the next token. But how do Midjourney, DALL-E, and Stable Diffusion create images? They use a completely different mechanism: diffusion.

Here's how diffusion models learn:

1. During training, the model sees millions of images paired with text descriptions.
2. For each image, the model watches a process of progressive noise addition: the image gets slightly corrupted with random static, then more corrupted, then more, until it's pure noise.
3. The model learns to recognize and reverse each step of this corruption — essentially learning to denoise an image guided by visual and textual patterns.

When you generate an image:

1. The model starts with pure random noise.
2. Your text prompt guides the denoising process — steering it toward visual patterns associated with your description.
3. Step by step, the noise becomes less random and more structured, until an image emerges that matches your prompt.

This is why more specific prompts produce more specific images — your text is actively steering the denoising process at every step, not just at the start.

This mechanism explains something remarkable: AI can generate images of things that never existed in the training data. No training image was labeled "an astronaut riding a horse on Mars at sunset in the style of Van Gogh." But the model learned visual concepts — astronauts, horses, Mars landscapes, Van Gogh brushwork — separately, and the diffusion process combines them under the guidance of your prompt.

Key contrast with language models: LLMs generate left-to-right, one token at a time. Diffusion models refine the entire image simultaneously across many denoising steps — the output of each step feeds into the next until the image converges.

Mapping the Broader Landscape

Multimodal models can accept multiple types of input — text, images, audio — and produce multiple types of output. GPT-4o, Gemini 1.5, and Claude 3 can all process images alongside text. Some can listen to audio and respond with voice. This convergence of modalities is a defining trend of 2025–2026.

Foundation models are large, general-purpose models trained on broad data that can be fine-tuned for specific tasks. ChatGPT is built on a foundation model. Most enterprise AI tools are fine-tuned versions of foundation models adapted for specific use cases: customer support, code review, document analysis.

Narrow AI refers to systems optimized for one specific task — a chess engine, an image classifier, a voice transcription system. These can exceed human performance on their specific task but cannot generalize beyond it. AlphaGo plays Go brilliantly; it cannot write an email.

Broad AI refers to systems that handle a wide range of tasks from a single model. Today's LLMs are broad AI — they can write, code, translate, summarize, and reason all within the same system.

Knowing what type of AI you're dealing with tells you what to expect. A narrow AI optimized for image classification isn't broken if it can't write; a broad AI assistant that fails at a specialized professional task might just need a better prompt or a more specific tool.