"AI" gets used as one giant catch-all word, which is a bit like calling every vehicle on the road "a car." Let's sort the traffic.
First, the Three Big Buckets
Almost everything people mean by "AI" falls into one of three nested categories - not separate universes, just increasingly specific circles:
- Machine learning - the broad family. Any system that learns patterns from data instead of being explicitly programmed rule-by-rule.
- Deep learning - machine learning's more ambitious sibling, using layered neural networks to learn much messier, higher-dimensional patterns. This is what made modern image and speech recognition actually work.
- Generative AI - deep learning models that don't just recognize patterns, they produce new content that fits them: text, images, video, audio, code, you name it. This is the category that ate the tech news cycle, and it's what the rest of this article is about.
So: generative AI is a type of deep learning, which is a type of machine learning.
Generative AI, By How It Actually Works
The "under the hood" view - the different mathematical approaches that all fall under "generative AI."
| Architecture | How it works, roughly | Great for | Example | Creator |
|---|---|---|---|---|
| GANs | Two networks compete - one fakes, one catches - until the fakes are convincing | Photorealistic faces, upscaling, style transfer | StyleGAN | NVIDIA Research |
| VAEs | Compresses data into a "concept space," then reconstructs variations from it | Smooth, controllable variation; often a building block inside bigger systems | The original VAE paper | Diederik Kingma |
| Diffusion Models | Starts from pure noise, gradually "denoises" it into something coherent, guided by your prompt | High-quality image and video generation - the technique behind most modern AI art | Stable Diffusion | Stability AI |
| Transformers / Autoregressive | Predicts the next piece of a sequence one step at a time, using self-attention to track context | Text, chat, code - and increasingly image/audio treated as sequences | GPT | OpenAI |
A fun twist: the lines have blurred hard. Modern image and video tools often mix diffusion and transformer architecture in the same pipeline. Nobody stays in their lane anymore.
Generative AI, By What It Actually Makes You
The practical view - organized by output, with a hosted app to try and an open-weight model to self-host if you'd rather keep it local.
| Type | Use it for | Try it (hosted) | Self-host instead |
|---|---|---|---|
| Text | Drafting, summarizing, brainstorming, coding help, arguing with yourself about email tone | ChatGPT, Claude, Gemini, Perplexity | Qwen 3 (see our sizing guide), Llama 3, Mistral, DeepSeek-V3, Gemma |
| Image | Concept art, marketing visuals, mockups, turning "a raccoon in a business suit" into a picture | Midjourney, DALL·E (via ChatGPT), Adobe Firefly, Ideogram | Stable Diffusion, FLUX (Black Forest Labs), SDXL, Playground v2 - run any of them with ComfyUI or Automatic1111 |
| Video | Short clips, ad concepts, storyboarding, previz without a camera crew | Runway, Pika, Kling, Luma Dream Machine | Wan (Alibaba), HunyuanVideo (Tencent), Mochi (Genmo), CogVideoX - all rougher than Runway, but yours |
| Music | Background tracks, jingles, full songs from a prompt on a deadline | Suno, Udio, ElevenLabs Music, Stable Audio | Stable Audio Open, MusicGen (Meta), AudioCraft - shorter, less polished, fully local |
| Voice | Narration, dubbing, cloning your own voice for a podcast intro | ElevenLabs, Play.ht, Murf, Speechify | Coqui XTTS, Bark, F5-TTS, OpenVoice - local voice cloning |
| Code | Autocomplete on steroids, boilerplate, explaining scary legacy functions | GitHub Copilot, Cursor, Replit AI, Codeium | Qwen 3 Coder, DeepSeek-Coder, Code Llama, StarCoder2 - open-weight and genuinely competitive |
| 3D | Turning a photo or prompt into a 3D asset for games, prototyping, or AR | Luma AI, Meshy, Spline AI, Tripo | TripoSR (Stability AI + Tripo), InstantMesh, Shap-E (OpenAI) - all run on a single consumer GPU |
The Honest Takeaway
None of this is permanent - the field reshuffles roughly every quarter, and tools merge, rebrand, or quietly get discontinued (a couple of video models didn't make this list because they vanished mid-draft). Treat it as a snapshot, not a bible. But the underlying shape holds up: machine learning learns patterns, deep learning learns messier patterns, and generative AI turns those learned patterns into something new. Everything else is just which flavor of math - and whether you're renting it or running it yourself.
