20 essential AI terms, explained in plain English. No PhD required.
An AI system that can take actions — not just answer questions, but browse, run code, or use tools to complete multi-step tasks.
How much text a model can 'remember' at once, measured in tokens. A 200K-token window fits roughly a 500-page book.
The technique behind Midjourney and DALL-E: start from random noise and gradually refine it into a coherent image.
A list of numbers representing the meaning of text. Similar meanings get similar numbers — it's how AI 'understands' similarity.
Taking a pre-trained model and training it further on specific data, so it specializes — e.g. a support bot trained on your help docs.
When an AI confidently states something false. It happens because models predict plausible text, not verified facts.
Running a trained model to get answers. 'Inference cost' is what you pay per query — distinct from the one-time training cost.
Large Language Model — AI trained on massive text datasets to understand and generate human language. GPT-6 and Claude are LLMs.
Low-Rank Adaptation — an efficient fine-tuning method that adjusts a small fraction of a model's weights, making customization cheap.
An architecture where a model contains specialized sub-models ('experts') and routes each input to the most relevant ones.
Handling more than text — images, audio, video. A multimodal model can look at a photo and describe it.
The input you give an AI. 'Prompt engineering' is the craft of writing inputs that get better outputs.
Shrinking a model by using less precise numbers, so it runs on smaller hardware with minimal quality loss.
Retrieval-Augmented Generation — connecting an LLM to your own documents, so answers are grounded in your data instead of hallucinations.
Reinforcement Learning from Human Feedback — training models to be helpful by having humans rank their outputs. Key to ChatGPT's quality.
Hidden instructions that set an AI's behavior — its role, tone, and rules — applied before your message.
A setting controlling randomness. Low temperature = predictable and focused; high = creative and surprising.
The unit AI models read text in — roughly ¾ of an English word. 'Hello world' is about 2-3 tokens. Pricing is usually per million tokens.
The neural network architecture behind all modern LLMs, introduced by Google in 2017. The 'T' in GPT.
Getting a good answer with no examples — just describing the task. Modern models are remarkably good at this.