Plain English explanations of the most important AI concepts — no technical background required.
AI (Artificial Intelligence) is software that can perform tasks that would normally require human intelligence — like understanding language, recognising images, or making decisions. Modern AI doesn't "think" the way humans do. Instead, it finds patterns in enormous amounts of data and uses those patterns to make predictions.
A Large Language Model (LLM) is a type of AI trained specifically on text. It reads trillions of words — books, websites, code, articles — and learns the patterns of language so well that it can generate new text that sounds human. When you type a question into ChatGPT or Claude, an LLM is reading your words and predicting, one word at a time, what a good response would look like.
It doesn't "know" things the way you know your own phone number. It has learned statistical patterns from a huge amount of text. When it seems to know a fact, it's really predicting what words are likely to follow your question, based on the patterns in its training data. This is why it can sometimes sound very confident while being completely wrong — a phenomenon called hallucination.
They are all LLM-powered chat assistants made by different companies — OpenAI makes ChatGPT, Anthropic makes Claude, and Google makes Gemini. They work in the same fundamental way but have been trained differently, with different data, different safety approaches, and different strengths. Claude tends to be strong at writing and nuanced reasoning. ChatGPT is widely used and fast. Gemini is deeply integrated with Google's products.
Most AI tools (ChatGPT, Claude) run on remote servers — you send your question over the internet, their computers process it, and send back an answer. "Running locally" means the AI model is downloaded and runs entirely on your own computer. Tools like Ollama make this possible. The advantage is privacy (nothing leaves your machine) and no ongoing cost. The limitation is that local models are typically smaller and less capable than the frontier cloud models.
AI models don't read word by word — they break text into "tokens," which are roughly word fragments. The word "unbelievable" might become three tokens: "un", "believ", "able". One token is roughly 0.75 words on average. This matters because models have a limit on how many tokens they can process at once (the "context window"), and API pricing is usually based on token count.
Hallucination is when an AI generates text that sounds plausible and confident but is factually wrong — it might invent a fake citation, misquote a statistic, or get a date wrong. It happens because the model is predicting likely-sounding text, not checking facts. You should always verify important facts from AI responses against reliable sources, especially for medical, legal, or financial matters.
This is genuinely debated. AI is already automating some tasks — routine writing, basic coding, data analysis, image generation. But most jobs involve complex judgement, relationships, physical presence, and creativity that AI handles poorly. The more likely near-term outcome is that people who know how to use AI effectively will be more productive than those who don't, rather than AI simply replacing roles wholesale.
Current AI tools have real risks: misinformation (AI-generated fake content), bias (models can reflect biases in their training data), misuse (fraud, spam, manipulation), and privacy risks. These are serious and worth understanding. The more speculative long-term risks — AI systems acting against human interests — are taken seriously by researchers at labs like Anthropic and DeepMind, and are an active area of safety research.
A curated collection of the best places to learn more about AI, large language models, and the broader landscape.
The 2017 Google research paper that introduced the Transformer architecture which underpins every modern LLM.
The clearest visual explanation of how transformers work. Essential reading for anyone wanting to understand the mechanics without heavy maths.
OpenAI's landmark 2020 paper introducing GPT-3 and the concept of few-shot learning at scale.
Free, hands-on course that takes you from basics to building real neural networks. No heavy maths prerequisites.
YouTube series by former Tesla and OpenAI researcher. Builds a neural network from scratch in Python. One of the best free resources available.
Free short courses on LLMs, prompt engineering, RAG, and agents. Taught by Andrew Ng and leading practitioners. Ideal for non-technical learners.
Weekly AI news and commentary from Andrew Ng. Clear, balanced, and accessible for non-technical readers.
Detailed weekly newsletter from the co-founder of Anthropic. Covers research, policy, and the strategic AI landscape.
Download and run open-weight models on your own computer. Free, private, no API costs.
The GitHub of AI models. Browse, download, and test thousands of open models. Essential reference for the open AI ecosystem.
LLMs 101 is an interactive mind map designed to help anyone — technical or not — build a solid mental model of how large language models work. It covers the mathematics, the training process, the major model architectures and families, and the prompting techniques that get the best results.
Each node in the map expands to reveal further detail. Clicking any node opens a sidebar with a full explanation and links to primary sources, so you can always go deeper.
Navigate to the Mind Map and start at the centre node — Large Language Models — and click it to reveal the four main branches. Click any branch to expand its children. Click any leaf node to open the detail sidebar. Use the hamburger menu (top left) to navigate between the different pages of this site.
On the mind map, nodes with a + indicator have children that will expand on click. Nodes without children open the detail panel directly.
Pure HTML, CSS, and JavaScript — no frameworks, no build tools, no dependencies beyond Google Fonts. Works as a single file that can be opened directly in any browser or hosted on any static web server.
Typography: Cormorant Garamond (headings) and Jost (body). Colour palette inspired by the Sahel theme by Qode Interactive — combining Dark Goldenrod, Tan, Dust Storm, White Smoke, and Van Dyke Brown.
All factual claims in the node explanations are drawn from primary sources including original research papers, official model documentation, and reputable technical publications. Sources are linked directly in each node's detail panel.
Key references include: Vaswani et al. 2017 (Transformer), Brown et al. 2020 (GPT-3), Ouyang et al. 2022 (InstructGPT/RLHF), and the official model cards from Anthropic, OpenAI, Meta, Google DeepMind, Mistral, and Alibaba Qwen.
If you notice that any of the content is incorrect, citations are missing or wrong, or if you have any suggestions for improvement, please feel free to reach out to us at: