This essential AI terms glossary exists because the field has a habit of minting new vocabulary faster than anyone can absorb it. Sit in on a product meeting these days and you’ll hear LLMs, RAG, RLHF and a dozen other abbreviations flying around, terms that can make even seasoned technologists feel slightly adrift.
Below are plain-language definitions of the terms you’re most likely to run into, whether you’re building, investing, or just trying to follow the conversation.
The Essential AI Terms Glossary, From LLMs to Agents
AGI, Artificial general intelligence is a phrase everyone uses and nobody quite agrees on. OpenAI’s charter defines it as ‘highly autonomous systems that outperform humans at most economically valuable work.’ Google DeepMind frames it as ‘AI that’s at least as capable as humans at most cognitive tasks.’ Even among the researchers closest to the problem, the definition shifts depending on who you ask.
AI agent, A system that carries out multi-step tasks on your behalf: filing expenses, booking restaurant tables, writing and maintaining code. The catch is that the infrastructure to deliver on those promises is still being built. ‘AI agent’ means different things to different people, partly because the field itself is still being defined.
Chain of thought, When a model breaks a complex problem into smaller intermediate steps before arriving at an answer, that’s chain-of-thought reasoning. It takes longer but produces more reliable results, particularly in logic or coding tasks. Reasoning models are built on this approach, using reinforcement learning to optimise for it.
Hallucination, The industry’s preferred euphemism for an AI model generating information that is simply incorrect. The problem is thought to arise from gaps in training data, and it’s a genuine obstacle: a health query that returns harmful medical advice is not a quirk, it’s a risk. The push towards more specialised, domain-specific AI models is partly a response to this.
Inference and training, Training is how a model learns: data goes in, patterns are identified, and the model adjusts its outputs toward a target. Inference is what happens when you actually use the model, applying what it learned to generate a response. Training is expensive and done once (or occasionally); inference happens every time someone sends a prompt.
Large language model (LLM), The engine behind ChatGPT, Claude, Gemini, Llama, Copilot and most other AI assistants you’ve used. LLMs are deep neural networks made of billions of numerical parameters, called weights, that encode the relationships between words and phrases drawn from billions of books, articles and transcripts. When you prompt one, the model generates the most statistically likely continuation of your input.
Model Context Protocol (MCP), Possibly the most consequential piece of AI plumbing most people have never heard of. Anthropic launched MCP in November 2024 as an open standard: a universal way to connect AI systems to external data sources (files, databases, apps like Slack or Google Drive) without developers having to build a custom connector for every possible pairing. The shorthand that stuck: USB-C for AI.
The protocol spread unusually fast. By the time Anthropic donated it to the Linux Foundation, MCP had accumulated more than 10,000 active servers and 97 million-plus monthly SDK downloads, according to Anthropic’s post on the donation.
On 9 December 2025, the Linux Foundation announced the formation of the Agentic AI Foundation (AAIF), with MCP as a founding project alongside Block’s ‘goose’ and OpenAI‘s ‘AGENTS.md’. Backing came from Anthropic, OpenAI, Block, Google, Microsoft, Amazon, Cloudflare and Bloomberg. Bloomberg’s stated aim, per the AAIF formation release, is to extend MCP so it can be used securely in regulated financial services environments. That framing tells you something about how seriously the enterprise world is taking this standard.
Reinforcement learning, A training method in which a model learns by attempting tasks and receiving a reward signal for correct outputs, rather than being trained on a fixed labelled dataset. Techniques like reinforcement learning from human feedback (RLHF) are now central to how leading AI labs refine their models to be more helpful and accurate.
Tokens and token throughput, Tokens are the chunks, often fractions of words, that language models break text into before processing. They’re also the unit of cost: most AI companies charge on a per-token basis. Token throughput measures how many tokens a system can process in a given period, which determines how many users a model can serve simultaneously and how quickly each receives a response.
Weights, Numerical parameters that determine how much importance a model assigns to different features in its training data. Training begins with randomly assigned weights; as the model learns, those weights adjust until outputs align with the target. In a housing-price model, weights encode how much factors like the number of bedrooms or the presence of parking actually influence a property’s value.
Why the Essential AI Terms Glossary Keeps Growing
The vocabulary expands because the technology expands. MCP going from a spec document to an eight-organisation foundation project in under a year is a useful illustration: niche plumbing becomes industry infrastructure almost overnight, and a term nobody knew becomes one you need to understand. That’s the pattern throughout this essential AI terms glossary, and it won’t slow down.
The next term likely to graduate from jargon to board-level concern: whatever framework ends up governing how AI agents are authorised to act on your behalf. The governance question is where the real friction is building, and the AAIF’s work on agentic standards is probably where to watch for it first.
