How Many Kinds of AI Are There? The Whole Map

When two people discuss “using AI”, they are often discussing different machines. One means a language model answering chat messages; the other means an image classifier running on a security camera. This chapter draws the map, so the rest of the series can name things precisely.

1.1 The nested layers

These terms are concentric circles. Each outer ring contains the ones inside it.

AI (artificial intelligence) is the widest umbrella — any system doing work that would normally take human intelligence. That includes if-else rule engines, pathfinding, and expert systems, none of which learn from data at all.

ML (machine learning) learns patterns from data instead of having a person write the rules. Decision Tree, Random Forest, XGBoost, SVM and Logistic Regression are all still in heavy use, and usually beat deep learning on tabular data.

DL (deep learning) is ML built on many-layered neural networks that learn their own features from raw input. It suits unstructured data: images, audio, text, video.

Foundation models are large DL models trained on a broad mass of data and then adapted to many downstream jobs. LLMs, VLMs and their relatives live at this layer — and this is what most people mean by “AI” in 2026.

The misunderstanding that costs the most money

Not every problem needs an LLM. For predicting numbers from tabular data — sales, loan defaults, customer churn — XGBoost is usually more accurate, a thousand times faster, explainable, and runs on a single CPU. Pointing an LLM at a job that classical ML does better means paying more for a worse result.

1.2 By how they learn

Method	Data it needs	Example jobs
Supervised	Inputs paired with correct answers (someone labelled them)	Spam classification, licence-plate reading, price prediction
Unsupervised	Raw data with no answers	Customer segmentation, anomaly detection, dimensionality reduction
Self-supervised	Raw data, with the task invented from it — mask a word, predict it	How every LLM is trained; the reason internet text works without labelling
Reinforcement	Learns from reward and penalty by trying	Games, robotics, and the RLHF stage that makes an LLM answer the way people want

The chat models in daily use pass through all three stages: self-supervised pre-training on an enormous body of text, then supervised fine-tuning on human-written conversations, then RL to shape what it agrees to say and what it refuses.

1.3 By architecture

Architecture	Era	Good at	Still used in 2026?
CNN	2012–	Images — spatial patterns	Very much so: object detection, first-pass OCR, embedded work that has to stay light
RNN / LSTM	2014–	Sequences such as text and audio	Nearly displaced by the Transformer; survives where compute is extremely tight
Transformer	2017–	Long sequences, every position attended at once	The backbone of almost everything today
Diffusion	2020–	Generating images, audio and video from noise	The standard for image generation
GAN	2014–	Image generation	Almost entirely replaced by diffusion; survives where speed is paramount
GNN	2018–	Graph data — molecules, social networks	Specialist work such as drug discovery and fraud detection
SSM (Mamba)	2023–	Very long sequences at linear cost	Still a second choice, usually seen hybridised with a Transformer

1.4 By what goes in and what comes out

This is the split you actually choose between when designing a system.

Short name	Full name	In → out	What it is for
LLM	Large Language Model	Text → text	Writing, summarising, translation, question answering, code, classification, extraction
VLM	Vision-Language Model	Image + text → text	Reading scanned documents, describing images, answering questions about charts, context-aware OCR
Omni	Multimodal	Image + audio + text → text/audio	Assistants that talk and watch a screen at the same time
VLA	Vision-Language-Action	Image + instruction → action	Robots, robotic arms, agents that drive a computer screen
ASR / STT	Speech Recognition	Audio → text	Meeting transcripts, subtitles, dictation
TTS	Text-to-Speech	Text → audio	Voiceover, voice assistants, voice cloning
Embedding	Text/Image Embedding	Text/image → numeric vector	Semantic search, RAG, clustering, similarity
Reranker	Cross-encoder	Query + document → score	Reordering search results for precision; usually paired with embeddings in RAG
T2I / T2V	Text-to-Image / Video	Text → image/video	Design, advertising, content
Guard	Safety Classifier	Text → safe/unsafe	Filtering input and output before it reaches a user

Real systems chain several of these

A receipt-reading system, for example: a VLM turns the image into structured text, a small LLM sanity-checks the numbers, an embedding model matches the shop name against a vendor database, and classical ML predicts the expense category. Splitting the work this way is both cheaper and more accurate than asking one large model to do all of it.

1.5 Terms people mix up

Term	What it means
Dense	Every parameter runs on every answer. Straightforward, but the bigger it gets the slower it is
MoE	Mixture of Experts — a large total parameter count of which only part activates per token, say 30B total with 3B active. Fast like a small model, but it consumes the VRAM of a large one, because all of it must be loaded
Reasoning model	Trained to think in steps before answering. Slower and more token-hungry, but more accurate on maths, logic and complicated code
Context window	The most a model can take in one pass, counted in tokens. The longer it is, the more VRAM it costs
Token	A fragment of text. English runs roughly one token per four characters; Thai costs far more, because most tokenizers were trained mainly on English
Fine-tuning	Continuing training from an existing model on your own data, to adjust style or add domain knowledge
LoRA / QLoRA	Fine-tuning that touches only a small slice of the parameters, which is what makes training a large model on a single card possible
RAG	Letting the model search your documents before answering — cheaper and faster than fine-tuning for most domain-knowledge problems
Distill	Using a large model to teach a small one, to get close capability at a size you can actually run
Quantization	Reducing the numeric precision inside a model to make it smaller and faster, at a small cost in quality

What this chapter settles

“AI” covers several layers, and matching the layer to the problem is the cheapest decision available on day one. Tabular data usually ends at classical ML; text work ends at an LLM; scanned documents need a VLM. And real systems tend to chain several models rather than lean on one.

The next chapter is about reading model names — what the numbers and suffixes at the end of a model name are actually telling you.