Every AI term
explained plainly
No jargon. No math. Just clear definitions of the AI terms you'll actually encounter — from RAG and LoRA to agentic AI, fine-tuning, and MCP. Updated weekly by our editorial team.
AI systems that can plan, execute, and self-correct over multi-step tasks autonomously, rather than just answering one question at a time. They use tools, browse the web, and iterate on goals.
A hypothetical AI that can perform any intellectual task a human can, across any domain. Current AI is "narrow" — it's superhuman at chess or coding but useless at unrelated tasks. AGI would generalize.
The mechanism that lets a transformer model "look at all words at once" and decide which ones matter for predicting the next one. It's why ChatGPT can resolve "it" in long sentences.
Running many prompts through a model at once instead of one at a time. Cheaper and faster per request, but with higher latency for any individual response.
A prompting technique where you ask the model to "think step by step" before answering. Dramatically improves accuracy on math, logic, and reasoning tasks.
The maximum amount of text (measured in tokens) a model can "see at once." Bigger = it can read more of your codebase, document, or chat history in a single call.
An agentic mode where the AI autonomously searches the web for 5–30 minutes, reads dozens of sources, and produces a long, cited report. Pioneered by OpenAI and Google.
The class of AI that powers image generators like Midjourney and Stable Diffusion. It learns to reverse the process of adding noise to an image, gradually "denoising" random pixels into a coherent picture.
A way of representing text (or images, audio) as a list of numbers that captures meaning. Two pieces of text with similar meaning will have similar embeddings.
Taking a pre-trained model and training it further on your own data so it learns your style, domain, or task better. Cheaper than training from scratch.
OpenAI's family of language models. "Generative" (it creates text), "Pre-trained" (on most of the internet), "Transformer" (the underlying architecture).
When an AI confidently states something false. The model isn't lying — it's generating plausible-sounding text based on patterns, and sometimes those patterns produce wrong facts.
When the model is running — generating output for a user query. Different from training, which is when the model learns. Inference is what costs money per API call.
A neural network trained on massive amounts of text to predict the next word. The category includes GPT, Claude, Gemini, Llama, etc. "Large" usually means 1B+ parameters.
A way to fine-tune a model by adding a small "adapter" instead of retraining the whole thing. 100× cheaper than full fine-tuning, with similar results.
An open standard from Anthropic that lets AI assistants plug into external tools (databases, APIs, file systems) in a uniform way. Think USB-C for AI agents.
A model that can handle more than one type of input — typically text, images, audio, and video together. GPT-4o, Gemini, and Claude are all multimodal.
The "knobs" inside a neural network that get tuned during training. More parameters = more capacity. GPT-3 has 175B parameters. Llama 4 has 405B.
The first, most expensive step of building an AI: feeding it huge amounts of text from the internet so it learns language patterns. Costs millions of dollars in GPU time.
The art of writing instructions for an AI so it gives you exactly what you want. Bad prompts → bad outputs. Good prompts → magic.
A technique that lets the AI look up information from a database or document collection before answering, instead of relying only on what it learned in training.
A class of model (e.g. OpenAI o1, DeepSeek R1) that "thinks" before responding, using extra compute at inference time to work through hard problems step by step.
The training step that aligns AI behavior with human preferences. Humans rate model responses; the model learns to produce responses humans prefer.
A setting (0–2) that controls how random the output is. Low = deterministic and focused. High = creative and varied. 0.7 is the typical default.
How much thinking time the model uses when answering. New reasoning models (o1, o3) get smarter by spending more compute at inference, not just at training time.
The basic unit of text an LLM processes. Roughly ¾ of a word in English. "ChatGPT is amazing" → 5 tokens. API pricing is per token.
When an AI is allowed to call external functions — search the web, run code, look up a database, send an email. Critical capability for agents.
The neural network architecture that powers every modern LLM. Introduced by Google in 2017 ("Attention Is All You Need"). Replaced earlier RNN-based models.
A specialised database for storing and searching embeddings. Lets you find "semantically similar" items quickly — the backbone of RAG, recommendation engines, and semantic search.
A term coined by Andrej Karpathy for building software by describing it in plain English to an AI agent, accepting whatever it writes without reading the code carefully.
A model that can see images and produce text. Lets you upload a photo and ask questions about it.
Asking the AI to do a task with no examples in the prompt. Modern LLMs are good at this — but few-shot (with examples) usually works better.