AI Is Not One Thing
TL;DR
'AI' covers at least four very different families of technology. Using them interchangeably makes us worse at thinking about any of them. Here is the map.
The word “AI” has quietly become useless.
It describes the spam filter in your inbox, the system that suggests your next video, the model that writes your emails, and the arm of a warehouse robot. These are not similar technologies. They work differently, fail differently, and matter differently. Collapsing them all into one word makes it harder to think clearly about any of them.
So here is the map, in four families.
1. Pattern Recognition
Given input, classify it.
Is this email spam? Is this mole malignant? Is this transaction fraudulent? Is this image a cat?
These systems don’t write, don’t speak, don’t reason. They recognize. Trained on labeled examples, they compare and categorize. Computer vision, medical imaging, fraud detection, language detection — all pattern recognition.
They are usually narrow, often accurate, and almost never what people mean when they say “AI is changing the world.”
2. Recommendation and Prediction
Given context, suggest what comes next.
Netflix’s next show. Spotify’s next song. Amazon’s next product. The feed on every social platform.
These systems watch what people do and guess what they will want. They don’t understand you. They pattern-match your behavior against millions of other people’s behavior, then reduce a vast space of options to one plausible next step.
This is where AI has had its largest measurable effect on human attention — and arguably, on human belief.
3. Generative Models
Given a prompt, create something new.
Language: ChatGPT, Claude, Gemini. Images: Midjourney, Stable Diffusion. Video: Sora, Runway. Music, code, voice.
These are the models behind the word “AI” in most public conversation right now. They produce. They combine. They mimic. They are the most dramatic and the most debated — because for the first time, the output looks like something a person made.
This is the family the last post was about.
4. Agents and Reinforcement Learning
Given a goal, act.
The robot arm that learns to stack. The game-playing system (AlphaGo, AlphaZero). The autonomous driving stack. The coding agent that opens a file, writes a patch, runs tests.
These systems don’t just classify, predict, or generate — they decide and act. They live inside a loop: try, observe the result, adjust, try again. The newer generative models are starting to be bolted into this loop, which is why “agent” is suddenly everywhere.
This is where many of the biggest risks and biggest opportunities of the next decade will sit.
Why This Matters
When someone says “AI will replace jobs,” the first question is: which kind?
Pattern recognition has been replacing certain jobs for twenty years. Recommendation systems changed the media industry, not the warehouse. Generative models affect writers and designers. Agents threaten routine workflows.
Each family has different strengths. Different failure modes. Different places they belong.
The most productive way to think about AI is not as one monolithic thing rolling toward us — it is as a set of tools, each with a specific shape, each asking a specific question:
Where does this actually fit in a life well-lived?
That is the only question worth asking.