Complete AI Guide 2026: How Modern AI Really Works

"Artificial intelligence" is one of the most stretched phrases in technology. It is used for spam filters, chess engines, recommendation feeds, chatbots that write essays and systems that predict the shape of proteins. Some of these are decades old, some are a few years old, and they work in quite different ways.

This guide is the hub of our AI coverage. It explains, in plain language, what AI is and is not, how machine learning and deep learning work, how large language models (LLMs) are built and trained, what agents and tool use add, how image and audio generators work, how AI systems are evaluated and where they fail, what safety and alignment research is trying to solve, what drives compute and cost, where AI is genuinely useful, and how to keep learning. Every section links to a deeper article on the topic.

We have deliberately avoided ranking "the best model right now". That kind of claim goes stale within weeks. The ideas below change much more slowly, and once you understand them you can judge new releases for yourself.

---

What AI Is, and What It Is Not

At its broadest, artificial intelligence is the field of building systems that perform tasks we would normally say require intelligence: recognising speech, translating text, planning a route, diagnosing a fault, writing a paragraph. The term itself dates to the 1956 Dartmouth summer research project, where John McCarthy and colleagues proposed studying how machines could "use language, form abstractions and concepts". Our history of artificial intelligence timeline walks through the long cycles of optimism and "AI winters" since then.

It helps to keep a few nested terms apart:

• Artificial intelligence (AI): the whole field, including old rule-based "expert systems" that encode human knowledge as if-then rules.
• Machine learning (ML): the subset of AI where a system learns patterns from examples instead of being given explicit rules. Almost all modern AI is ML.
• Deep learning: the subset of ML that uses neural networks with many layers.
• Generative AI: models that produce new content (text, images, audio, video, code) rather than only labelling inputs. Our explainer on what generative AI is and how it is changing industries covers the business side.
• Large language models (LLMs): deep learning models trained on very large text collections to predict and generate language.
• Artificial general intelligence (AGI): a hypothetical system that matches human ability across essentially all cognitive tasks. There is no agreed definition and no agreed test, which is one reason forecasts vary so wildly. See what AGI means and why experts disagree.

What AI is not is also worth stating. Current systems are generally understood to have no goals of their own beyond their training objective and the task they are given, do not "look things up" unless they are connected to a search tool, and have no built-in reliable way to check whether what they say is true (researchers debate how much a model internally represents about truth). An LLM produces text that is statistically likely given its training and the prompt. That is often correct and useful, and sometimes confidently wrong. Keeping that mental model in place prevents most of the common mistakes people make with these tools.

---

How Machine Learning Works

Traditional software is a set of instructions written by a programmer. Machine learning flips this: you supply examples, and an algorithm adjusts a model's internal numbers (its parameters) until the model's outputs match the examples well. Our beginner-friendly guide, what machine learning is, explained simply with real examples, is the best starting point if this is new to you.

There are three classic styles of learning:

1. Supervised learning: every training example has a correct answer (an email labelled spam or not spam, a house with its sale price). The model learns to map inputs to answers.
2. Unsupervised learning: there are no labels, and the model looks for structure on its own, such as grouping customers with similar behaviour. How clustering finds hidden groups shows how this works.
3. Reinforcement learning: an agent takes actions, receives rewards or penalties, and learns a strategy that maximises reward over time. Game-playing systems and the later stages of chatbot training both use it.

The core loop: loss and gradient descent

Nearly every modern model is trained the same way at heart. The model makes a prediction, a loss function measures how wrong it was, and an optimisation method nudges each parameter slightly in the direction that reduces the loss. That method is almost always some form of gradient descent, explained step by step in how gradient descent finds better answers. Repeat this over millions or billions of examples and the model gradually improves.

Classic algorithms still matter

Not every problem needs a neural network. For tabular business data (rows and columns), decision trees and their ensembles are often faster, cheaper and easier to explain. Read how decision trees split data, then how random forests reduce error and why gradient-boosted trees such as XGBoost keep winning Kaggle competitions. For a single tour of regression, k-nearest neighbours, support vector machines and more, see ML algorithms explained.

Generalisation: the real goal

A model that memorises its training data is useless on new data. That failure is called overfitting, and the opposite failure, a model too simple to capture the pattern, is underfitting. The standard defence is to hold back data the model never sees during training and score it there. Our articles on the overfitting and underfitting trade-off and why training and test data must be separated explain this discipline, which applies just as much to LLM evaluation as to a simple regression.

Data is most of the work

Practitioners often say that the model is the easy part. Collecting, cleaning, labelling and understanding data takes most of the time in real projects. What data cleaning actually involves and what feature engineering means describe the unglamorous work that decides whether a model is any good. Statistical literacy matters too: a model that finds a strong pattern has found a correlation, not necessarily a cause, as how correlation differs from causation explains.

---

Deep Learning and Neural Networks

A neural network is a stack of layers. Each layer takes a list of numbers, multiplies them by learned weights, adds them up, applies a simple non-linear function, and passes the result on. With enough layers and enough data, such networks can learn remarkably rich representations: edges, then shapes, then objects in images; characters, then words, then meaning in text. How neural networks work builds this up from a single artificial neuron.

The modern deep learning era is usually dated to 2012, when the AlexNet network (Krizhevsky, Sutskever and Hinton) won the ImageNet image recognition challenge by a large margin using a deep convolutional network trained on graphics processors (GPUs). That result showed that big networks, big datasets and parallel hardware together could beat hand-engineered methods.

The Transformer

The architecture behind today's language models is the Transformer, introduced by Google researchers in the 2017 paper "Attention Is All You Need" (Vaswani et al., arXiv:1706.03762). Its key idea is self-attention: when processing each word (strictly, each token, a word or word fragment), the model computes how relevant every other token in the input is and blends information accordingly. Unlike earlier recurrent networks, Transformers process a whole sequence in parallel, which made it practical to train them on far more data.

For an intuitive walkthrough of attention, read how Transformers pay attention to words; for the full architecture, see Transformers explained simply. If you want to build models yourself, the choice of framework matters less than people think, but PyTorch vs TensorFlow vs JAX lays out the trade-offs.

---

How Large Language Models Are Built and Trained

An LLM is a Transformer trained to do one deceptively simple thing: given a sequence of tokens, predict the next one. Everything a chatbot does, from answering questions to writing code, emerges from that objective plus additional training. How does ChatGPT actually work? is our plain-language companion to this section, and how AI models are trained, step by step covers the full pipeline in more depth.

Stage 1: Pretraining

The model reads an enormous corpus of text and code and repeatedly guesses the next token. Each wrong guess produces a small parameter update. Over trillions of tokens, the model absorbs grammar, facts, styles of reasoning and the structure of code, because all of those help predict text.

How big should a model be for a given compute budget? DeepMind's 2022 "Chinchilla" paper (Hoffmann et al., arXiv:2203.15556) found that many large models of the time were undertrained, and that model size and the amount of training data should be scaled up roughly in proportion. That finding shaped how later models were built and is a good example of the durable research results that outlast any particular model release.

Some training data is now generated by other models rather than written by people. Synthetic data explained covers why labs do this and what can go wrong.

Stage 2: Instruction tuning

A pretrained model is a powerful autocomplete, not an assistant. Ask it a question and it may continue with more questions. Supervised fine-tuning trains it further on curated examples of instructions paired with good responses, teaching it the format of helpful dialogue.

Stage 3: Learning from feedback

The third stage uses preferences. In reinforcement learning from human feedback (RLHF), people compare pairs of model outputs, a separate reward model learns to predict their preferences, and the main model is optimised to score well. OpenAI's InstructGPT paper (Ouyang et al., 2022, arXiv:2203.02155) reported that labelers preferred outputs from a 1.3-billion-parameter instruction-tuned model over those of the much larger 175-billion-parameter GPT-3. Newer variants replace some human judgments with AI judgments (RLAIF) or skip the separate reward model (DPO); our comparison of RLHF, RLAIF and DPO explains the differences, and how AI models learn from feedback gives the general picture.

Inference: what happens when you type a prompt

At use time ("inference"), your text is split into tokens, run through the network, and the model outputs a probability for every possible next token. A sampling rule picks one, it is appended, and the process repeats. Settings such as temperature control how adventurous that choice is, which is why the same question can produce different answers; temperature and top-p explained covers the knobs.

The model can only attend to a limited span of tokens at once, its context window. Anything outside it is invisible. Context windows explained covers what that limit means in practice.

Reasoning models and test-time compute

A more recent development is training models to work through a problem in intermediate steps before answering, spending more computation at inference time on harder questions. This "test-time compute" approach improved performance on maths and coding tasks. Test-time compute and LLM reasoning explains the mechanism, and how LLMs actually reason looks at what that "reasoning" is and is not. DeepSeek's openly released R1 model, covered in DeepSeek R1 explained, made one version of this technique widely inspectable.

Architecture variations

Two ideas come up repeatedly in model announcements. Mixture of experts models contain many sub-networks and activate only a few for each token, which lets total parameter count grow without a proportional rise in per-token cost; see how mixture-of-experts routing works. Distillation trains a smaller "student" model to imitate a larger "teacher", producing cheaper models that keep much of the capability; see AI model distillation explained.

---

Before a model sees any text, it is split into tokens, which explains several odd failures such as miscounting letters; see how LLM tokenization works. How loss falls as models, data and compute grow is covered in neural scaling laws.

Grounding Models in Real Information: Embeddings and RAG

Because an LLM's knowledge is frozen at training time and stored diffusely in its weights, many real systems give it relevant documents at query time instead. This is retrieval-augmented generation (RAG), a term from a 2020 paper by Lewis et al. at Facebook AI Research (arXiv:2005.11401).

RAG relies on embeddings: numerical vectors that place texts with similar meaning close together. A question is embedded, the most similar passages are found in a store of document embeddings, and those passages are pasted into the prompt. Embeddings and vector search explained covers the maths in plain terms, and what RAG is and how it works covers the full pattern, including why retrieval quality usually matters more than which model you use. For how assistants combine context, retrieval and stored notes, see AI memory systems explained.

RAG reduces but does not eliminate made-up answers: if the retrieved passage is wrong or missing, the model can still produce a confident answer.

---

Agents and Tool Use

A plain LLM only produces text. Tool use (also called function calling) lets the model output a structured request, such as "call the weather API with city = Paris", which the surrounding software executes and feeds back. Strictly defined outputs make this reliable; see function calling vs tool use.

An agent is a loop around this: the model decides on a step, a tool runs, the result comes back, and the model decides the next step until the task is done. AI agents explained is our main deep dive, and tool use vs agents draws the line between a single call and an autonomous loop.

Standards are emerging for connecting models to tools. Anthropic announced the Model Context Protocol (MCP) on 25 November 2024 as an open standard for linking AI assistants to data sources and tools; it has since been adopted by other providers. Read what MCP is, and for the business-software angle, how AI agents connect to business apps.

Agents fail in characteristic ways: errors compound over many steps, they can misread a web page, and they can be manipulated by text planted in the content they read. That last problem, prompt injection, is serious enough to have its own article: prompt injection explained. For a realistic assessment, read what AI agents can and cannot do.

Agents also overlap with classic automation. A workflow tool that runs fixed steps on a trigger is predictable; an agent that chooses its own steps is flexible but less predictable. Pairing the two, with deterministic workflows for routine steps and approval workflows that catch automation mistakes before anything irreversible happens, is a common design. If you want to try this, the n8n workflow automation beginner's guide is a practical place to start.

---

Generative Image, Audio and Video Models

Most current image generators are diffusion models. During training, the model sees images progressively corrupted with random noise and learns to reverse each small step. At generation time it starts from pure noise and denoises step by step, guided by a text prompt. The approach was popularised by "Denoising Diffusion Probabilistic Models" (Ho, Jain and Abbeel, 2020, arXiv:2006.11239). Our explainer on how DALL·E, Midjourney and Stable Diffusion work goes deeper, and how image generators understand prompts covers the link between text encoders and pictures.

Video generators extend the same ideas across time, which is far more compute-hungry and makes consistency between frames hard; see what Sora is. Audio and music models generate waveforms or compressed audio tokens; AI music generation tools and AI voice tools survey the practical options. Speech recognition, the reverse direction, has its own failure modes, covered in why AI transcription tools make different mistakes.

Generative media raises distinct issues: copyright in training data, consent for voice cloning, and synthetic images used to deceive. These are covered in the ethics discussion in the safety section below.

---

The Main Branches of AI at a Glance

Subfield	What it does	Everyday example	Start here
Classical machine learning	Learns patterns from structured data	Credit scoring, churn prediction	Random forest explained
Recommender systems	Ranks items a user is likely to want	Video and shopping feeds	How recommendation systems rank choices
Time-series forecasting	Predicts future values from past data	Demand and energy forecasts	How time series forecasting works
Natural language processing / LLMs	Understands and generates text	Chat assistants, translation	Prompt engineering explained
Computer vision	Interprets images and video	Photo search, medical imaging	AI image editing tools
Generative media	Creates images, audio, video	Illustration, voiceover	Free AI image generators
Agents	Plans and acts with tools	Coding agents, research assistants	What agentic AI is
Scientific AI	Models physical and biological systems	Protein structure prediction	How AI is changing drug discovery

(The "start here" column points to a practical companion piece; the conceptual deep dives are linked in the sections above and below.)

---

Evaluation: How Do We Know a Model Is Any Good?

Measuring AI quality is harder than it looks. Public benchmarks (fixed sets of questions with known answers) are useful for tracking progress, but they have well-known weaknesses: test questions can leak into training data, models can be tuned to the benchmark format, and a single score hides wide variation across tasks. Why AI benchmarks can be misleading explains these problems, and why data leakage breaks AI evaluations covers the contamination issue specifically.

For anyone deploying a model, the more useful practice is building evals: your own test sets that reflect your actual tasks, scored automatically or by people, and rerun whenever you change the prompt or model. LLM evaluation explained shows how. Once live, systems need monitoring as well; AI observability covers tracing and quality checks in production, and what model drift means explains why a model that worked at launch can degrade as the world changes. For an independent annual view of the field, the Stanford HAI AI Index report compiles benchmark, investment and policy data each year.

When judging a claim, ordinary statistics still applies: small test sets give noisy results (why sample size changes confidence), and a reported improvement may fall within the margin of error.

---

Limitations: Hallucination and Bias

Hallucination

LLMs sometimes state false things fluently: invented citations, wrong dates, non-existent API functions. This is called hallucination. It follows from how the models work. They are trained to produce plausible continuations, and a plausible-sounding falsehood can score well. Training methods that reward guessing over admitting uncertainty can make it worse. Why AI hallucinates facts covers causes and mitigations: grounding with retrieval, asking for sources and checking them, lowering sampling randomness for factual tasks, and keeping a human in the loop for anything that matters.

Bias

Models learn from human-generated data, and they absorb the patterns in it, including stereotypes and the under-representation of some groups and languages. Bias can also come from how a problem is framed or which data was easy to collect. Data-science fundamentals help here: survivorship bias is one example of how the data you have can mislead you about the world. Bias cannot be removed by a single fix; it requires testing outcomes across groups, documenting limitations and deciding where automated decisions are appropriate at all.

Other practical limits

• Knowledge cut-off: without retrieval, a model knows nothing after its training data ends.
• Brittleness: small changes in wording can change answers.
• Context limits: very long inputs may be truncated or attended to unevenly.
• Security: prompt injection and data leakage risks grow as models gain tools and access.

---

Safety and Alignment

Alignment is the problem of getting AI systems to do what their designers and users actually intend, including in situations nobody anticipated. For today's systems that means refusing clearly harmful requests, being honest about uncertainty, and not taking unwanted actions when operating as agents. Researchers also study longer-term risks from more capable systems, such as models pursuing proxy goals that diverge from intended ones.

Practical techniques include the feedback-based training described above; Anthropic's "Constitutional AI" approach (Bai et al., 2022, arXiv:2212.08073), in which a model critiques and revises its own outputs against a written set of principles; red-teaming, where people deliberately try to elicit harmful behaviour; and interpretability research, which tries to understand what is happening inside a network.

Governance has moved from discussion to formal rules and frameworks. The US National Institute of Standards and Technology published its voluntary AI Risk Management Framework (AI RMF 1.0) on 26 January 2023. The European Union's AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024, with obligations phasing in over several years: bans on certain practices applied from 2 February 2025 and rules for general-purpose AI models from 2 August 2025. Our articles on the key debates in AI ethics and the future of artificial intelligence discuss where these debates are heading.

---

Researchers are also trying to read what happens inside these models; mechanistic interpretability describes the methods and how far they reach.

Compute and Cost

Training a frontier model requires large clusters of specialised chips (GPUs or similar accelerators), running for weeks or months, plus the energy and cooling to support them. Groups such as Epoch AI track the training compute of notable models and have documented rapid growth over the past decade. Because these figures and costs change quickly, we do not quote specific numbers here; Epoch AI's public database and the Stanford AI Index are the best places to check current estimates.

For users and developers, costs come mainly from inference, usually priced per token. Several techniques reduce it:

• Smaller models handle many routine tasks well. See what small language models are good for.
• Quantization stores weights at lower precision so models fit on cheaper hardware; see LLM quantization explained.
• Prompt caching avoids reprocessing the same long instructions on every request; see prompt caching explained.
• Model routing sends easy queries to cheap models and hard ones to expensive ones; see AI model routing explained.
• Efficient serving engines batch many requests together on the same hardware; see what vLLM is.

Running models on your own computer is increasingly practical with tools such as Ollama (what Ollama is), but it has its own trade-offs in hardware, speed and quality, set out in the hidden cost of running a local LLM on a laptop. For subscription costs, our AI chatbot pricing comparison is dated and updated separately, and why free AI tools are not always cheaper is worth reading before you choose.

---

Real-World Uses That Hold Up

Beyond the demos, a few areas have clear, documented value:

• Software development. Code assistants autocomplete, explain and refactor code, and agentic tools can make multi-file changes. They still need review. See what Cursor is and what Claude Code is, and for the risks of accepting code you have not read, what vibe coding is.
• Science. DeepMind's AlphaFold predicts protein structures from amino-acid sequences; Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for it (with David Baker, for computational protein design). See AlphaFold and protein folding.
• Healthcare. Imaging analysis, triage support and documentation are active areas, with strong requirements for validation and oversight; see AI in healthcare.
• Search and research. Assistants that search the web and cite sources change how people find information, though citation practices vary; see AI search vs traditional search and deep research AI tools.
• Everyday knowledge work. Summarising, drafting and analysing documents. Our guides to using AI to automate daily tasks, using AI for studying and AI data analysis tools cover what works.

On jobs, the honest answer is that AI changes tasks more often than it removes whole occupations, and the effects differ widely by field. How AI will change jobs by 2030 looks at the evidence and its limits.

---

Using AI Well: Practical Habits

A few habits separate useful AI use from frustrating use:

1. Give context and constraints. Say who the output is for, what format you need and what to avoid. How to write AI prompts that actually work has worked examples.
2. Ask for structure when software will read the output. JSON mode vs structured outputs explains the options.
3. Verify specifics. Treat names, numbers, quotes and citations as unconfirmed until checked.
4. Keep sensitive data out of tools whose data policies you have not read.
5. Judge tools on your own tasks, not on marketing. How to evaluate AI writing tools without hype gives a method that works for other tool categories too.

---

How to Keep Learning

AI changes fast at the surface but slowly underneath. A sensible path:

1. Foundations first: basic Python, some linear algebra and probability, and the ideas in this guide. The complete machine learning learning path lays out an order.
2. Take a structured course. The best machine learning courses compares well-known options.
3. Build small projects with real data, and evaluate them honestly.
4. Learn deployment if you want to work in industry; MLOps for beginners is the bridge from notebook to production.
5. Read primary sources. Papers on arXiv, official model documentation and system cards, the Stanford AI Index and Epoch AI's data are more reliable than social media summaries.

For career planning, the ML engineer career path covers skills and roles.

---

FAQ

Q: Is AI the same as machine learning? A: No. Machine learning is a subset of AI in which systems learn from data. Older AI included rule-based systems written by hand. In practice, almost all AI products today are built with machine learning, which is why the terms are often used interchangeably.

Q: How does a chatbot like ChatGPT or Claude produce answers? A: It is a Transformer-based language model that predicts text one token at a time. It was pretrained on large text collections, then fine-tuned on instruction examples and human or AI feedback so its responses are helpful and follow instructions.

Q: Why do AI models make things up? A: They are trained to produce plausible text, not to check facts, so a fluent but false answer can look like a good continuation. Retrieval of trusted documents, checking sources and human review reduce the problem but do not remove it.

Q: What is the difference between an AI assistant and an AI agent? A: An assistant responds to each message. An agent runs in a loop, choosing steps, calling tools such as search or code execution, and using the results to decide what to do next until a task is finished. Agents are more capable but make compounding errors and face prompt-injection risks.

Q: Does artificial general intelligence exist yet? A: There is no agreed definition or test for AGI, and no current system is widely accepted as meeting it. Current models are strong on many language and coding tasks but remain unreliable in ways a general human-level intelligence would not be.

Q: Is AI regulated? A: Increasingly. The EU AI Act entered into force on 1 August 2024 with obligations phasing in over several years, and NIST's voluntary AI Risk Management Framework, published in January 2023, is widely used in the United States. Rules vary by country and sector.

Q: What is the best way to start learning AI? A: Use AI tools regularly on real tasks, learn the core ideas of machine learning and LLMs, then follow a structured course and build small projects. Read primary sources rather than relying on summaries.

---

Bottom Line

Modern AI is mostly machine learning, and its most visible form, the large language model, is a Transformer trained to predict text and then shaped by human feedback. That design explains both its strengths (fluent language, broad knowledge, useful reasoning on many tasks) and its weaknesses (hallucination, inherited bias, sensitivity to wording). Understanding the mechanics lets you use these tools where they help, verify them where they can mislead, and read new announcements with appropriate scepticism.

This hub is reviewed as the field evolves. Last reviewed: October 2026.