Terms, A to Z
20 terms
AGI
Artificial general intelligence: AI that matches people across most thinking tasks. There is no agreed test for it.
AI agent
An AI that takes actions toward a goal, such as clicking, filling in forms or running code, instead of only replying. Grok Bot is an agent.
AI data center
A building full of GPUs, power and cooling, used to train and run AI models. Colossus is xAI's.
API
A way for software to call a model directly, instead of through a chat app. Grok's is documented at docs.x.ai.
Context window
How much text a model can consider at once, measured in tokens.
Gigawatt (GW)
A billion watts. The size of an AI data center is now often quoted in power, not chip count.
GPU
A GPU (graphics processing unit) is a chip first built to speed up graphics. Its strength at parallel math also makes it the main chip for training and running AI models, which is why AI clusters are counted in GPUs.
HBM (high-bandwidth memory)
Stacked memory placed next to an AI chip to feed it data fast. Its supply is often the bottleneck in making AI chips.
Inference
Inference is running an already-trained AI model to produce an answer, such as a chatbot reply. It is separate from training, which builds the model, and it runs for every user request.
Inter-satellite laser links
Lasers that let Starlink satellites pass data to each other in space.
Large language model (LLM)
An AI model trained on huge amounts of text to predict the next word, which lets it answer questions, write and code. Grok is a family of LLMs.
Low Earth orbit (LEO)
Orbits up to about 2,000 km above Earth. Starlink's satellites fly there.
Multimodal
A model that handles more than text, such as images, audio or video.
Orbital compute
Running AI computers on satellites instead of on the ground, powered by sunlight. Musk's memo at the February 2, 2026 xAI acquisition described it as a main aim of the merger.
Parameters
The adjustable numbers inside a model that training sets; a rough measure of the model's size.
Reasoning model
A model that works through intermediate steps before answering, trading speed for accuracy on hard problems.
Reusable booster
A rocket's first stage that lands and flies again, a large part of why Falcon 9 can fly so often.
Superintelligence
Superintelligence means an intelligence far beyond human ability. Philosopher Nick Bostrom's 2014 book Superintelligence: Paths, Dangers, Strategies (Oxford University Press) popularized the term and discussed how such a system might be controlled.
Token
A chunk of text, often part of a word, that models read and write. API prices are quoted per million tokens.
Training compute (FLOP)
Training compute is the total count of floating-point operations (FLOP) used to train an AI model. It is a count of work done, unlike FLOPS, which measures operations per second.
People
Who's who
The people whose public statements show up on the timeline, with what they said and when.
FAQ
Explained questions
What is the difference between training and inference?
Training builds a model by running huge amounts of data through it, once, at great cost. Inference is using the finished model to answer each request. Training is a project; inference is a running bill.
What is orbital compute?
Running AI computers on satellites powered by sunlight instead of on the ground. Our orbital compute page explains it, with the dated source for each claim.
What does superintelligence mean?
An intelligence far beyond human ability in almost every area. Nick Bostrom's 2014 book Superintelligence made the term widely known.
Answers as of November 4, 2026.