Glossary

Each term is defined once here, in everyday words. Pages across the site link to these definitions the first time a term appears.

Actuator: A part that makes a robot move, usually an electric motor with gears.
AI agent: An AI system that can take several steps on its own, such as using tools, to reach a goal.
Alignment: The challenge of making an AI system pursue the goals its builders and users actually intend.
Artificial intelligence (AI): Computer systems that do tasks which usually need human judgment, such as recognizing pictures or writing text.
Attention: A technique that lets a model decide which earlier words matter most when it picks the next one.
Chatbot: A product you talk to in everyday language. It usually runs a language model underneath.
Context window: The amount of text a model can look at at one time, like its working memory for a conversation.
Controller: Fast software that turns a target (such as 'bend this joint 20 degrees') into precise motor commands.
Deep learning: Machine learning that uses neural networks with many layers.
Embodied AI: AI that senses and acts in the physical world through a body, such as a robot.
Feedback training (RLHF): Training where people (or other AI) rate answers, and the model is nudged toward the answers that get better ratings.
Fine-tuning: Extra training on a smaller, chosen set of examples to make a model better at a specific job or style.
Hallucination: When an AI gives a fluent, confident answer that is false or invented.
Large language model (LLM): A model trained on huge amounts of text to predict what text should come next.
Machine learning: A way of building software by showing it many examples so it finds patterns, instead of writing every rule by hand.
Model: The trained program itself: a big set of weights plus the code that runs them.
Moravec's paradox: The observation that things humans find hard (like chess) can be easy for computers, while things toddlers do easily (like walking) can be very hard.
Neural network: A program made of many simple steps arranged in layers, each controlled by adjustable numbers called weights.
Open weights: A model whose trained weights can be downloaded. This is not always the same as fully open source.
Pretraining: The first, biggest training stage, where a model learns patterns from a very large collection of data.
Sensor: A part that measures something about the world or the robot itself, such as light, distance, or joint angle.
Teleoperation: A person controlling a robot remotely instead of the robot acting on its own.
Token: A small chunk of text (a word, part of a word, or punctuation) that a language model reads and writes one piece at a time.
Training: The process of adjusting a model's weights using many examples until its answers get better.
Transformer: The neural-network design, introduced in 2017, that most modern language models are built on.
Vision-language-action (VLA) model: A model that takes in camera images and a written instruction and outputs robot actions.
Weights (parameters): The adjustable numbers inside a neural network. Training changes them. Think of them as dials.