Artificial Intelligence, Explained From First Principles

"Artificial intelligence" is a marketing term that covers everything from a spam filter to a system that writes working software. Strip the branding away and almost all of it reduces to one idea: a program whose behaviour is learned from data rather than written by hand. Understanding that idea — what is learned, how, and what can go wrong — is enough to reason clearly about most of what's happening in the field.

What an AI system actually does

A classical program encodes rules: if the email contains these words, flag it. A machine-learning model instead has millions or billions of adjustable numbers (parameters) and a training procedure that nudges them until the model's outputs on example inputs match the desired outputs. Once trained, it's just a function: input in, prediction out. The prediction might be a label ("spam"), a number (tomorrow's demand), a ranking (which video next), or — in generative models — a whole new artefact such as a paragraph, an image or a block of code.

The reason this matters is that the rules are never written down. They exist implicitly in the parameters, which is why models can capture patterns no human could specify, and also why they can capture patterns nobody wanted.

How models learn

Three broad paradigms cover most of practice:

• Supervised learning uses labelled examples — images tagged with what they contain, transactions marked fraudulent or not. Most deployed business ML is still this.
• Self-supervised learning creates its own labels from raw data: hide a word and predict it, hide a patch of an image and reconstruct it. This is how large language and vision models are pretrained on internet-scale data without anyone labelling it.
• Reinforcement learning learns from reward signals rather than correct answers. It trains game-playing agents and robots, and — as reinforcement learning from human feedback (RLHF) and its successors — it's how chat models are shaped to be helpful rather than merely fluent.

Deep learning is the sub-field where the model is a neural network with many layers. Its defining advantage is that it learns useful intermediate representations on its own: the early layers of a vision model discover edges and textures, the later ones objects, with no one designing those features. The transformer architecture, introduced in 2017, made this scale to language and is now the backbone of nearly every frontier model, in text, images, audio and code.

Perception, reasoning, action

A useful way to think about AI systems is as a pipeline of three capabilities. Perception turns raw signals into structure: transcribing speech, detecting objects, parsing a document. Reasoning works over that structure to plan or decide: which route, which diagnosis to consider, which tool to call next. Action produces an effect in the world: sending a message, moving a robot arm, executing a trade. Older systems did one of these. Modern agents chain all three in a loop, observing the result of each action and re-planning — which is why they can complete multi-step tasks and also why their errors can compound.

Running AI in production

Operating a model is mostly ordinary distributed-systems work with some extra dials. You still watch latency, error rates and cost. But you also watch for things that don't exist in conventional software:

• Data drift. The world changes; the model's training snapshot doesn't. A fraud model trained before a new payment method launched will quietly degrade.
• Feedback loops. A recommender shapes what users see, which shapes the data it's retrained on. Left alone, that loop narrows.
• Offline/online gap. A model that scores well on a held-out test set can still fail in production because the test set wasn't representative of real traffic.
• Non-determinism. Generative models can give different answers to the same prompt. Testing them needs evaluation suites and statistical thresholds, not exact-match asserts.

Mature teams treat models as living artefacts: versioned, evaluated on a schedule, deployed behind canaries and shadow tests, with a clear path to roll back. This discipline has a name — MLOps — and it's where most of the practical difficulty lives.

Safety and ethics are design decisions

Bias, privacy and misuse aren't bolt-ons; they're properties of choices made early. A model trained on historical hiring data learns historical hiring bias. A model trained on user messages can memorise and regurgitate them. Techniques exist for each — balanced datasets and fairness metrics, federated learning and differential privacy, output filtering and red-teaming — but they only help if they're specified as requirements rather than discovered as incidents. Regulation is catching up: the EU AI Act's obligations for high-risk systems are phasing in through 2026–27, and NIST's AI Risk Management Framework has become the reference vocabulary in the US. AI governance is now a job title.

Where the field is heading

Three trends are visible from here. Models are becoming multimodal by default, reasoning jointly over text, images, audio and video. They are becoming agentic, given tools and autonomy to complete tasks rather than answer questions. And, in parallel with ever-larger frontier models, small specialised models are proliferating — distilled, quantised and run on laptops, phones and edge devices where latency, cost and privacy make the cloud the wrong place for inference.