Artificial intelligence is usually classified in three ways: by capability (narrow, general and super AI), by functionality (reactive, limited memory, theory of mind and self-aware), and by technique (rule-based systems, machine learning, deep learning and generative models). Every AI system in use today, including large language models, is narrow, limited-memory AI.
Why are there different ways to classify AI?
Different people ask different questions. A policymaker asks "how capable could this become?", a researcher asks "what kind of memory and reasoning does it have?", and an engineer asks "how is it built?". Each question gives a different classification, and all three are useful when you are learning the field.
What are the types of AI by capability?
| Type | Meaning | Exists today? | Examples |
|---|---|---|---|
| Narrow AI (weak AI) | Performs specific tasks, often better than people | Yes: all current AI | Spam filters, face unlock, translation, LLM chatbots |
| Artificial general intelligence (AGI) | Human-level ability across almost any intellectual task | No | Research goal of several labs |
| Artificial superintelligence (ASI) | Exceeds the best human ability in every domain | No (hypothetical) | Discussed in AI safety research |
Modern LLMs blur the line a little because they handle many tasks (writing, coding, summarising, translating) with one model. They are still considered narrow: they make basic reasoning errors, can't reliably learn new skills on their own after training, and have no grounded experience of the world.
What are the types of AI by functionality?
This classification, popularised by researcher Arend Hintze, describes how a system relates to memory and other minds:
- Reactive machines respond to the current input with no memory of the past. IBM's Deep Blue, which beat Garry Kasparov at chess in 1997, is the classic example.
- Limited memory systems use past data to make decisions. Every trained machine learning model qualifies: it learned from historical data, and a chatbot also uses the conversation so far as short-term memory.
- Theory of mind AI would understand that others have beliefs, intentions and emotions. Not achieved.
- Self-aware AI would have consciousness and a sense of self. Purely hypothetical.
What are the types of AI by technique?
For anyone learning to build AI, this is the most useful view, because each technique is a skill you can learn:
| Technique | How it works | Strengths | Weaknesses |
|---|---|---|---|
| Rule-based / expert systems | Humans encode if-then rules | Transparent, predictable | Brittle, hard to scale |
| Classical machine learning | Algorithms learn patterns from tabular data | Fast, explainable, strong on structured data | Needs feature engineering |
| Deep learning | Multi-layer neural networks learn features themselves | Images, audio, text at scale | Needs more data and compute |
| Reinforcement learning | Agent learns from rewards through trial and error | Sequential decisions, games, control | Hard to train, sample-hungry |
| Generative models (LLMs, diffusion) | Learn the distribution of data and create new samples | Text, code and image generation | Hallucination, cost, evaluation |
Real products mix these. A bank's fraud system may combine rules, gradient-boosted trees and a neural network; an AI agent combines an LLM with rule-based guardrails.
How do the learning approaches differ?
Within machine learning, there are three main ways a model learns:
- Supervised learning: learns from labelled examples (house features → price). Most business ML is supervised.
- Unsupervised learning: finds structure in unlabelled data, such as customer segments with k-means clustering.
- Reinforcement learning: learns from rewards; also used to align LLMs with human preferences.
LLMs add a fourth idea, self-supervised learning: the model learns by predicting the next word in huge amounts of text, so the data labels itself.
Which types should you learn first?
Bar length is proportional to the week in the 78-week programme. Source: Program Zero curriculum.
Start with classical machine learning: it teaches training, validation and overfitting with models you can fully understand. Then deep learning, then transformers and LLMs. Reinforcement learning appears where it matters for LLMs (alignment with PPO and DPO) in the advanced phase.
Common misconceptions about types of AI
- "ChatGPT is AGI." It is a capable narrow system; it still fails at tasks a child could do and cannot verify its own facts.
- "Deep learning replaced classical ML." For tabular business data, gradient-boosted trees are often still the best choice. See XGBoost explained.
- "Rule-based AI is obsolete." Rules remain essential for safety limits, compliance and guardrails around ML systems.
How do these types show up in a single product?
Take a familiar example: a food-delivery app. Within one app you can find almost every type of AI described above, working together:
- Rule-based logic decides whether a restaurant is open, whether an address is serviceable and whether a coupon applies.
- Classical machine learning (often gradient-boosted trees) estimates delivery time from distance, traffic, kitchen load and weather.
- Deep learning reads menu photos or matches dish images to listings.
- Unsupervised learning groups customers with similar tastes for recommendations.
- An LLM powers the support chat that answers "where is my order?" and summarises complaints for human agents.
- Optimisation and reinforcement-style methods help assign delivery partners to orders.
Every one of these is narrow, limited-memory AI. None of them "understands" food or cities. Yet combined with good engineering, they make a product that feels intelligent. That is the reality of AI work today: picking the right technique for each part of a problem and connecting them reliably.
How do you choose the right type of AI for a problem?
| If your problem is… | Start with | Why |
|---|---|---|
| Clear, stable logic with few exceptions | Rules | Cheapest, fully explainable |
| Predictions from tables of numbers and categories | Classical ML (trees, regression) | Strong accuracy with modest data and compute |
| Images, audio or long text | Deep learning or a pretrained model | Learns features humans can't hand-craft |
| Open-ended language tasks | An LLM with retrieval and guardrails | Flexible, but needs evaluation |
| Sequential decisions with feedback | Reinforcement learning or optimisation | Learns policies over time |
A good rule: always try the simplest approach that could work first, and measure it. Fancy techniques earn their place only when they beat a simple baseline on data the model has never seen.
Will the types of AI change?
The capability categories (narrow, general, super) will stay useful as long as AGI remains an open question. The technique categories change faster. Ten years ago, most practical AI was classical machine learning; today generative models dominate the headlines, and agents that plan and use tools are the active frontier. What does not change is the foundation underneath: programming, data, statistics and the core ideas of learning from examples. Learn those well and new types of AI become new tools rather than new subjects to start from scratch.
Which careers map to which types of AI?
Each technique has its own kind of practitioner. Data scientists and ML engineers mostly work with classical ML and deep learning on business data. Computer-vision and speech engineers specialise in deep learning for images and audio. LLM and AI engineers build with generative models, retrieval and agents. Research engineers work on training methods, scaling and alignment. Almost all of them share the same base: Python, data handling, statistics and an understanding of how models learn. That shared base is why a broad, ordered programme prepares you for any of these paths rather than locking you into one early.
Where to go next
If you want to build these systems rather than just name them, start with what AI is and how it works, then follow a structured path. Program Zero's AI course covers classical ML, deep learning, NLP, LLMs and agents in order, and the machine learning course page details the algorithms you will train yourself.