What is AI?

Q: "What is the difference between AI and machine learning?"

"AI (Artificial Intelligence) is the broad concept: software that performs tasks normally requiring human intelligence. Machine learning (ML) is the main technique used to achieve modern AI: instead of programming explicit rules, you train a system on data and it learns the rules itself. Deep learning is a subset of ML using neural networks with many layers. In practice: when people say 'AI' in 2026, they usually mean ML-based systems, often deep learning. When they say 'machine learning', they might mean traditional ML (decision trees, regression) or neural networks depending on context."

Q: "What is a token in AI?"

"Language models process text as tokens, pieces of words, not always whole words. 'unbelievable' might be split into 'un', 'believ', 'able'. The exact split depends on the tokeniser. Models have a context window measured in tokens: how much text they can consider at once. GPT-4o handles 128,000 tokens (~100,000 words). Claude models support up to 200,000 tokens. The limit matters because everything outside the context window is not 'visible' to the model when it generates a response."

Q: "Why does AI make things up (hallucinate)?"

"Language models generate text by predicting what token comes next, based on patterns in their training data. They have no mechanism for distinguishing 'I know this' from 'I am extrapolating this'. When asked about something outside their training data or at the edges of their knowledge, they continue generating plausible-sounding text, which may be factually wrong. This is hallucination: fluent, confident, incorrect output. Mitigation strategies: retrieval-augmented generation (RAG), grounding responses with source documents, and always verifying factual claims from AI independently."

AI is software that learns patterns from data instead of following hand-written rules. Here is what that actually means, without the hype.

7 min read No prior knowledge needed

Recommended watch

But What is a Neural Network?

3Blue1Brown

Watch on YouTube →

Quick Answer

AI (Artificial Intelligence) is a broad term for software that performs tasks normally requiring human intelligence. Modern AI works by learning statistical patterns from enormous amounts of data, rather than following explicit rules written by a programmer. Large language models like Claude, GPT, and Gemini are the dominant form of AI in 2026.

The old way vs the new way

Traditional software follows explicit rules written by programmers:

python

if age < 18:
    deny_purchase()
if temperature > 100:
    send_alert()

Every situation must be anticipated and coded manually. This works for precise, rule-governed tasks. It breaks down for complex ones, recognising a cat in a photo, understanding sarcasm, or generating a coherent essay.

Machine learning takes a different approach: instead of writing rules, you give the system thousands or millions of labelled examples, and it figures out the patterns itself. Show it enough photos labelled “cat” and “not cat” and it learns what makes a cat a cat, better than you could describe in rules.

The result is systems that can do things programmers could never explicitly code. More: Machine learning, Google Developers

Traditional software

Developer writes explicit rules

Every case must be anticipated in code

Deterministic: same input always gives same output

Easy to explain: "if X then Y"

Breaks on edge cases not in the rules

Good for: accounting, form validation, business rules

Machine learning

System learns rules from labelled examples

Generalises to cases it has never seen

Probabilistic: outputs have confidence scores

Hard to explain: parameters are opaque weights

Degrades gracefully on novel inputs

Good for: image recognition, language, recommendations

A black geometric prism with a precise red laser entering one face and refracting through the structure: raw input transformed into structured, weighted output. — Traditional code is explicit rules written by hand. Machine learning is a system that learns the rules from data. The same input goes in. The transformation is different. The output is calibrated, not declared.

Types of machine learning

Supervised learning, the most common type. You provide labelled training examples (inputs paired with correct outputs). The model learns to predict the output for new inputs. Examples: spam detection, image classification, sentiment analysis, house price prediction.

Unsupervised learning, you provide inputs without labels. The model finds structure on its own. Examples: customer segmentation, anomaly detection, topic modelling.

Reinforcement learning, the model learns by taking actions in an environment and receiving rewards or penalties. Examples: game-playing AI (AlphaGo, AlphaStar), robotic control, recommendation systems. RLHF (Reinforcement Learning from Human Feedback) is used to align language models with human preferences.

Self-supervised learning, the model creates its own labels from the data structure. This is how language models are trained: predict the next word given all previous words. Doing this on billions of documents develops rich internal representations of language and knowledge.

What is a large language model (LLM)?

An LLM (Large Language Model) is the type of AI behind Claude, ChatGPT, Gemini, and similar tools. It is a neural network trained on a vast amount of text, books, websites, code, conversations, scientific papers, using self-supervised learning.

The training objective is simple: given a sequence of tokens, predict the next token. Doing this well at scale (on hundreds of billions of parameters, across trillions of tokens) requires the model to develop deep representations of language, knowledge, and reasoning.

When you send a message to Claude, the model does not “look up” the answer. It generates a response one token at a time, each token being the model’s prediction of what should come next given all the preceding context.

Key LLMs and who builds them:

Claude, Anthropic. Model docs
GPT-4o, OpenAI
Gemini, Google DeepMind
Llama 3, Meta (open weights, downloadable)
Mistral, Mistral AI (European, open source)
Command R, Cohere (enterprise-focused)

Neural networks: the architecture

An LLM is built on a neural network, loosely inspired by the brain, practically a mathematical function. 3Blue1Brown’s video series on neural networks is the clearest visual explanation.

The core components:

Parameters (weights): numerical values that the model learns during training. GPT-3 had 175 billion parameters. Parameter counts for frontier models are no longer publicly disclosed; the industry trend since 2023 has moved toward smaller, more efficient models (7B-70B range) for most applications.
Layers: stacked transformations applied to the input. Each layer learns to recognise increasingly abstract patterns.
Attention mechanism (Transformer): the architectural innovation from 2017 that enabled modern LLMs. Attention lets the model consider relationships between distant parts of the input, understanding that “it” in “The cat sat on the mat. It was comfortable.” refers to the cat. More: Attention is All You Need, original paper (readable abstract)

Training vs inference

Training builds the model. Billions of text examples, gradient descent updating parameters over weeks on thousands of GPUs. This is enormously expensive, frontier model training costs tens to hundreds of millions of dollars.

Inference runs the model. Given your prompt, generate a response. This is much cheaper and is what happens every time you send a message.

Most people and applications use inference via APIs. Only a handful of labs train frontier models from scratch.

Fine-tuning is a middle ground: take a pre-trained model and train it further on a smaller specialised dataset. This adapts a general model for a specific domain (medical, legal, coding) at a fraction of the cost of training from scratch.

AI lifecycle: where you fit in

Pre-training

Trillions of tokens Thousands of GPUs Weeks → months Done by AI labs (Anthropic, OpenAI, Google). Cost: $50M–$500M+

Fine-tuning

Domain-specific data Hours → days Done by companies adapting a base model to their use case

Inference

Your prompt API call Milliseconds What happens every time you use Claude, ChatGPT, or any LLM product

What AI is good at

Task	Examples
Text generation	Writing, summarising, translating, explaining, coding
Reasoning and analysis	Structuring problems, drafting plans, reviewing documents
Code generation	Writing, debugging, refactoring, explaining code
Image understanding	Describing images, answering questions about visual content
Image generation	DALL-E, Midjourney, Stable Diffusion, Adobe Firefly
Speech recognition	OpenAI Whisper, Google Speech-to-Text
Structured output	Extracting data from documents, classifying text

What AI is bad at

Hallucination, confidently generating false information. Always verify factual claims.
Real-time knowledge, models have a training cutoff date and do not know what happened since (without retrieval tools).
Precise arithmetic, LLMs are surprisingly poor at multi-step calculation. Use a calculator via tool use.
Counting and spatial reasoning, many models count tokens, not discrete objects.
Consistent long-form reasoning without tools, complex multi-step reasoning degrades over many steps without scaffolding.
Knowing what they don’t know, models often cannot accurately report their own uncertainty.

Context windows and RAG

The context window is how much text a model can consider at once. Everything outside the window is invisible to the model. For long documents, codebases, or multi-session conversations, this matters.

RAG (Retrieval-Augmented Generation) is the standard solution: embed and index your documents, retrieve relevant chunks at query time, and inject them into the context alongside the question. This lets a model answer questions about documents far larger than its context window, and ground answers in up-to-date sources. More: RAG, Anthropic docs

AI in production vs AI as a tool

There is a difference between using an AI tool (ChatGPT, Midjourney) and building something with AI APIs. When you build:

Your code calls an AI API (Anthropic, OpenAI, Google) with a prompt
The model generates a response
Your code uses that response: displays it, extracts data from it, triggers another action based on it

This is how AI-powered products work. The AI is one component inside a larger system, alongside a database, a server, APIs, and user interface. The earlier articles in this series describe all those other components.

A human silhouette facing a vast red-lit industrial system: one person, many components, all integrated into a single production system. — A production AI system is not just an LLM. The model is one component alongside retrieval, databases, business logic, and a user interface. The architect sees the whole. Understanding all the other components in this series is what lets you build and evaluate these systems.

What’s next

Next: What is Vibe Coding? , how to use AI to build software without needing to code yourself.

Official documentation: Anthropic Documentation

Frequently asked questions

What is the difference between AI and machine learning?

AI (Artificial Intelligence) is the broad concept: software that performs tasks normally requiring human intelligence. Machine learning (ML) is the main technique used to achieve modern AI: instead of programming explicit rules, you train a system on data and it learns the rules itself. Deep learning is a subset of ML using neural networks with many layers. In practice: when people say ‘AI’ in 2026, they usually mean ML-based systems, often deep learning. When they say ‘machine learning’, they might mean traditional ML (decision trees, regression) or neural networks depending on context.

What is a token in AI?

Language models process text as tokens, pieces of words, not always whole words. ‘unbelievable’ might be split into ‘un’, ‘believ’, ‘able’. The exact split depends on the tokeniser. Models have a context window measured in tokens: how much text they can consider at once. GPT-4o handles 128,000 tokens (~100,000 words). Claude models support up to 200,000 tokens. The limit matters because everything outside the context window is not ‘visible’ to the model when it generates a response.

Why does AI make things up (hallucinate)?

Language models generate text by predicting what token comes next, based on patterns in their training data. They have no mechanism for distinguishing ‘I know this’ from ‘I am extrapolating this’. When asked about something outside their training data or at the edges of their knowledge, they continue generating plausible-sounding text, which may be factually wrong. This is hallucination: fluent, confident, incorrect output. Mitigation strategies: retrieval-augmented generation (RAG), grounding responses with source documents, and always verifying factual claims from AI independently.