This section gives you the technical grounding to understand what is actually happening when you use these tools. It matters, because understanding the architecture helps you understand the limitations.

Before we get into the technical details, the single most important thing to understand is that AI in 2026 is not one thing. It is not ChatGPT. It is not the chatbot your company bolted onto its intranet. It is not the image generator your nephew used to make a funny picture of the dog. It is an ecosystem of dozens of tools built on different technologies, each good at different things. ChatGPT is one of the most well-known, but it is not always the best, and for many tasks it is not even the right category of tool. Understanding this landscape is what this guide is for.

Large Language Models (LLMs)

The tools at the centre of this guide, Claude, ChatGPT, Gemini, Perplexity, are all powered by Large Language Models. An LLM is a neural network (specifically a transformer architecture, introduced by Google researchers in 2017) that has been trained on enormous volumes of text: books, websites, code repositories, academic papers, and more.

Training works in two phases. First, the model learns to predict the next word in a sequence by processing billions of text examples (pre-training). This gives it a statistical understanding of language, facts, reasoning patterns, and code. Second, it is fine-tuned using human feedback (RLHF, reinforcement learning from human feedback, or similar techniques) to make it helpful, safe, and conversational rather than just a raw text predictor.

When you send a prompt to Claude or ChatGPT, the model generates a response token by token (a token is roughly three-quarters of a word), predicting the most likely next token given everything that came before it. It does not "think" in the way humans do. It does not have a database it looks things up in. It generates text based on patterns learned during training. This is why it can produce fluent, coherent, and often remarkably insightful responses, and also why it can confidently generate complete nonsense. It is always predicting plausible text, whether or not that text is factually correct.

Context window is a term you will encounter frequently. It refers to how much text the model can consider at once: your prompt, any uploaded documents, and the conversation history. Context windows have grown dramatically. Early models handled about 4,000 tokens (roughly 3,000 words). Current frontier models like Claude Opus 4.6 and Gemini 3 Pro handle 1 million tokens (roughly 750,000 words). A larger context window means the model can work with bigger documents without losing track of details.

Key models as of March 2026: Anthropic's Claude (Opus 4.6, Sonnet 4.6), OpenAI's GPT-5 series and o3/o4 reasoning models, Google's Gemini 3 Pro and Flash, Meta's Llama (open source), Mistral (open source, French), Alibaba's Qwen (open source), and DeepSeek (open source, Chinese). The commercial models (Claude, ChatGPT, Gemini) are the most capable. The open-source models are catching up fast and can be run locally on your own hardware (covered in Part Five).

Reasoning Models

A recent development worth understanding: reasoning models. OpenAI's o3 and o4-mini, and Claude's extended thinking mode, use a technique where the model explicitly reasons step by step before generating its final answer. Rather than immediately producing output, the model generates a chain of thought (often hidden from you), working through the problem methodically.

This significantly improves performance on complex tasks: mathematics, formal logic, multi-step coding problems, scientific reasoning, and strategic analysis. The trade-off is speed and cost. Reasoning models are slower and consume more resources. For everyday tasks, they are overkill. For hard problems, they are genuinely better.

Beyond LLMs: Other Types of AI in This Guide

Not everything in this guide is powered by an LLM. Several other AI architectures are at work:

Image and video generation

Diffusion models power image generation (Midjourney, DALL-E, Ideogram, Adobe Firefly) and video generation (Runway, Kling, Sora). These models are trained on image-text pairs and learn to generate images by starting with random noise and progressively refining it into a coherent image guided by your text description. This is fundamentally different from how LLMs work. It is why image generators can produce photorealistic visuals but struggle with text in images (they are manipulating pixels, not understanding language).

Speech synthesis and voice cloning (ElevenLabs) use models trained on audio data to generate natural-sounding speech. Voice cloning models learn the characteristics of a specific voice from a short sample and can then generate new speech in that voice. These are specialised neural networks, not LLMs, though they may use LLM components for text understanding.

Automatic speech recognition (ASR) powers transcription tools like Otter.ai and Fireflies. These models convert audio to text using architectures trained on vast amounts of spoken language paired with transcripts. OpenAI's Whisper model, which is open source, has been particularly influential here.

Music generation (Suno, Udio) uses specialised models that understand musical structure, genre conventions, vocal styles, and production techniques. These combine multiple AI approaches including audio generation and language understanding (for lyrics).

Computer vision is what enables the camera-based features discussed in the Camera section later in this guide. When you photograph something and ask Claude or ChatGPT "what is this?", the model processes the image through a vision encoder that converts visual information into a representation the LLM can reason about. Modern frontier models are natively multimodal, meaning they process text and images (and increasingly audio and video) within a single architecture.

Agents: The Next Architectural Shift

The latest development is the move from models that respond to models that act. Tools like Claude's Cowork, Manus, and OpenClaw represent a shift toward agentic AI: systems that can plan a sequence of actions, execute them (browsing the web, managing files, calling APIs, sending messages), observe the results, and adjust their approach. These agent systems typically use an LLM as the "brain" but wrap it in a framework that gives it access to tools and the ability to take actions in the real world.

This is still early. Agent systems are powerful but unreliable. They hallucinate, loop, consume resources unpredictably, and sometimes take actions you did not intend. Understanding that agents are LLMs with tool access (not a fundamentally different kind of intelligence) helps set realistic expectations about what they can and cannot do.

What AI Is Not

AI in its current form is not sentient, conscious, or genuinely "intelligent" in the way humans are. It does not understand what it is saying. It does not have beliefs, intentions, or experiences. It is very sophisticated pattern matching on an enormous scale. This is not a limitation that will be fixed in the next version. It is a fundamental characteristic of the architecture.

This matters practically: AI will confidently present false information because it is optimising for plausible text, not truth. It will agree with you when it should push back, because agreement is a common pattern in its training data. It will produce impressive-sounding but vacuous output if you give it vague instructions, because vague inputs produce generic patterns.

The people who get the most value from AI are the ones who understand this. They verify facts. They give precise instructions. They treat AI as a powerful tool with known failure modes, not as an oracle.