“A good map doesn’t explain everything — but it keeps you from getting lost.”
Executive Summary
There are 15 distinct AI types worth knowing, organized across four axes: what AI can do (capability), how it learns, how it’s built, and what job it performs. This blog gives an overview of their definitions, how they compare, and examples of each type of AI.
Learning method: how it was trained
Architecture: how it’s built
Function: what job it does
Capability
How much can it do?
1
ASI
Artificial Specialized Intelligence
Capability
What it is
Artificial Specialized Intelligence is extraordinarily good at one specific thing. However, that deep specialization actually makes it worse at everything else since it is built so far in one direction that it cannot function outside that lane at all.
Examples
Chess engines
Medical image diagnosis
Protein folding models
The real talk
AlphaGo can beat every human alive at Go but it would fail at tic-tac-toe if you changed the rules slightly. Its specialization isn’t a side effect, it’s the whole design. Most AI in production today fits here, even when the marketing doesn’t say so.
2
Ag
General AI (AGI)
Capability
What it is
An AI that can learn and do anything a human can. It isn’t limited to one task: it would be able to switch from writing a poem to fixing a car engine to diagnosing a disease, the way people can. This does not exist yet.
Examples
Long-term research goal
Theoretical benchmarks
The real talk
Nobody agrees on exactly what AGI would look like, when it might arrive, or whether we’re close. Take a look at 2026 AGI predictions by some of the most prominent figures in the industry here. But the honest answer is: we don’t know.
3
Sup
Superintelligence
Capability
What it is
An AI smarter than the best human minds at everything. Science, art, problem-solving, all of it at once. This is purely theoretical. Nothing like it exists.
Examples
Nick Bostrom’s research
The real talk
This is the scenario that AI safety researchers worry about most. Not because it’s coming tomorrow, but because if it did arrive and we weren’t prepared, it could be very hard to manage.
Learning Method
How was it trained?
4
Sv
Supervised Learning
Learning method
What it is
The AI is trained using examples that have already been labeled with the right answer. So, it studies thousands of examples until it can recognize the pattern on its own.
Examples
Spam detection
Sentiment analysis
Medical diagnosis tools
The real talk
This is the most common way AI is trained today. The big catch: if the labeled examples are wrong or biased, the AI learns those mistakes too. Therefore, what goes in shapes what comes out.
5
Us
Unsupervised Learning
Learning method
What it is
The AI looks for patterns in data without being told what to look for. There are no labels and no right answers, it just finds structure on its own.
Examples
Customer segmentation
Anomaly detection
Grouping similar things
The real talk
This is useful when you don’t know what categories exist yet. The AI discovers them. The tricky part is that it’s hard to check if the answer is actually correct, since there’s no answer key.
6
Rl
Reinforcement Learning
Learning method
What it is
The AI learns by trying things and getting feedback, like a reward when it does something right and a penalty when it doesn’t. It keeps adjusting until it gets better at earning rewards.
Examples
Robot movement
Game-playing AI
Training chatbots to be helpful
The real talk
This is part of how Claude was trained to give useful responses instead of weird ones. The catch: the AI learns to earn rewards, not to do what’s actually good. Those two things aren’t always the same.
7
Ss
Self-supervised Learning
Learning method
What it is
The AI creates its own practice problems from raw data. For example, it covers up a word and tries to guess it, then checks itself. No human has to label anything.
Examples
BERT
Most modern AI assistants
The real talk
This is the big reason AI took off so fast recently. Training on huge amounts of internet text requires no expensive human labeling. Bigger data plus bigger models led to abilities nobody fully expected. Researchers are still figuring out exactly why.
Architecture
How is it built?
8
Nn
Neural Networks
Architecture
What it is
A system made of layers of connected nodes, loosely inspired by how the brain works. Each layer finds patterns in the data and passes them to the next layer.
Examples
Image-recognition models
Voice recognition
Diffusion models
The real talk
A specific type called the Transformer took over the entire AI field starting around 2017. It works by letting every part of the input pay attention to every other part at the same time. Simple in theory, but wildly powerful in practice.
Sy
Symbolic / Rule-based
Architecture
What it is
This type uses hand-written rules and logic to make decisions. Basically, “if this happens, then do that.” You can read the rules and understand exactly why it made a choice.
Examples
Legal decision tools
Medical checklists
Decision trees
The real talk
People called this approach outdated when deep learning took off. Then, they realized you can’t explain a neural network’s decision in court. Rule-based AI is still widely used anywhere you need to show your work, e.g. medicine, law, finance.
10
By
Bayesian Systems
Architecture
What it is
This type makes decisions based on probability. It starts with a best guess, then updates that guess as new information comes in. It’s always asking: how confident am I, and what changes my confidence?
Examples
Medical diagnosis tools
Risk assessment
The real talk
This approach is honest about uncertainty in a way most AI systems aren’t. It forces you to say how sure you are before you decide. It’s harder to scale up to large systems, but the core idea is making a comeback in new forms.
11
Ev
Evolutionary Algorithms
Architecture
What it is
This type finds solutions the same way evolution works. It starts with many random options, keeps the best ones, mixes and mutates them, and repeats. Over many rounds, a good solution emerges.
Examples
Scheduling problems
Engineering design
The real talk
It’s slow compared to other methods, but useful for problems where there’s no clear path to the answer. It’s often working quietly inside tools that no one labels as “evolutionary AI” in the marketing.
Function
What job does it do?
12
Gn
Generative AI
Function
What it is
Generative AI creates new content — text, images, video, audio, code — by learning patterns from existing content and using those patterns to make something new.
Examples
Midjourney
Sora
Stable Diffusion
GitHub Copilot
The real talk
“Generative AI” describes what the AI does, not how it’s built. Most people use the term to mean all AI right now. That’s not accurate but it’s understandable, since this is the type most people actually interact with.
13
Pr
Predictive AI
Function
What it is
Predictive AI looks at past data to forecast what will happen next. It finds patterns in history and uses them to make a prediction.
Examples
Weather forecasting
Predicting customer behavior
Credit scoring
The real talk
This type has been quietly running large parts of the economy for over twenty years. Banks, insurance companies, and retailers all use it. So, most of the AI revolution that people think is new actually already happened here, in the background, in a spreadsheet.
14
Ag2
Agentic AI
Function
What it is
Agentic AI takes action in the real world on its own. It can browse the web, write and run code, send emails, and manage files without needing a human to guide each step.
Examples
Research agents
Computer-use tools
AutoGPT
The real talk
This is the fastest-moving area in AI right now. When AI goes from answering questions to actually doing things, the stakes change. However, a mistake isn’t just a wrong answer, but it might be a deleted file or a sent email. That’s why people are paying close attention to how safe these systems are.
15
An
Analytical AI
Function
What it is
Analytical AI digs through large amounts of data to find patterns and insights. The goal is understanding what the data means, not predicting the future or creating new content.
Examples
Recommendation engines
Detecting unusual activity
Search ranking
The real talk
This is the most common type of AI in big companies, even if they don’t advertise it that way. The recommendations you get on streaming services and social media? That’s analytical AI deciding what to show you next. It quietly shapes a huge amount of what you see online every day.
Quick Reference
| # | Sym | Name | Axis | One-liner | Examples |
|---|---|---|---|---|---|
| 1 | ASI | Artificial Specialized Intelligence | Capability | Extraordinary at one specific task, useless outside of it. | AlphaGo, Chess engines, Medical imaging |
| 2 | Ag | General AI (AGI) | Capability | Can learn and do anything a human can. Does not exist yet. | Theoretical, research goal |
| 3 | Sup | Superintelligence | Capability | Smarter than the best human at everything. Purely theoretical. | Theoretical only |
| 4 | Sv | Supervised Learning | Learning | Trained on labeled examples with known right answers. | Image classifiers, Spam filters |
| 5 | Us | Unsupervised Learning | Learning | Finds patterns in data with no labels or right answers given. | Topic modeling, Customer segmentation |
| 6 | Rl | Reinforcement Learning | Learning | Learns through trial and error. Rewards for good, penalties for bad. | AlphaGo, Game-playing AI |
| 7 | Ss | Self-supervised Learning | Learning | Creates its own practice problems from raw data. No human labels needed. | GPT models, BERT, Modern AI assistants |
| 8 | Nn | Neural Networks | Architecture | Layered connected nodes, loosely inspired by the brain. | Transformers, Image recognition |
| 9 | Sy | Symbolic / Rule-based | Architecture | Explicit hand-written rules. You can read exactly why it decided. | Expert systems, Decision trees |
| 10 | By | Bayesian Systems | Architecture | Probability-based. Updates its confidence as new information arrives. | Spam filters, Risk assessment |
| 11 | Ev | Evolutionary Algorithms | Architecture | Mimics natural selection: random options, keep the best, repeat. | Engineering design, Scheduling |
| 12 | Gn | Generative AI | Function | Creates new content — text, images, video, audio, code. | Claude, Midjourney, Sora |
| 13 | Pr | Predictive AI | Function | Uses past data to forecast what happens next. | Fraud detection, Credit scoring |
| 14 | Ag2 | Agentic AI | Function | Takes real-world action autonomously. | Claude Code, Research agents |
| 15 | An | Analytical AI | Function | Finds patterns and meaning in large amounts of data. | Recommendation engines, Search ranking |