GO DEEPER / 03
Glossary
THE SHORT ANSWER
Short definitions of the terms our explainers use in AI, robotics, space and engineering failures. Each definition comes from the article it links to, where the sources are listed, so you can check it. Terms are sorted alphabetically.
A
- A* search
- A path-finding method that looks for the cheapest route using an estimate of the distance still to go. It was developed for the Shakey robot at SRI. Read more: Shakey, the Robot That Gave Us A*
- Actuator
- The part that makes a joint move, usually a motor combined with gearing. Read more: How Humanoid Robots Actually Work
- AI agent
- A language model that chooses its own steps, calls tools and checks the results in a loop until a task is done. Read more: What Is an AI Agent, Really?
- Alignment tax
- The loss of performance on some tasks that can come with making a model follow human preferences. OpenAI's InstructGPT paper reduced it by mixing the original pre-training objective back in. Read more: What RLHF Does and Does Not Do
B
- Benchmark
- A fixed set of tasks used to compare systems. Benchmarks age: once they leak into training data or models saturate them, they stop telling systems apart. Read more: What Is an AI Agent, Really?
C
- Colocation
- A data centre where several companies rent space, power, cooling and connectivity for their own hardware. Read more: Data Centres: Where the Power and Water Go
- Convex optimisation
- A class of maths problems with no false bottoms: a solver cannot get stuck on an answer worse than the best one, and it can be shown to finish. This makes it suitable for on-board landing guidance. Read more: How Rockets Land Themselves
- Core rope memory
- The fixed memory of the Apollo Guidance Computer, in which software was woven into wires: a bit was stored by a wire. Read more: The Apollo Guidance Computer
- CubeSat
- A small satellite built to a standard size, based on a 10 cm cube called 1U. Larger ones are built from several units, such as 3U or 6U. Read more: Why CubeSats Changed Access to Space
D
- Dispenser
- The box between a CubeSat and its rocket. It holds the satellite during launch and releases it into orbit when the launch vehicle sends the signal. Read more: Why CubeSats Changed Access to Space
- Dispersion
- The variation in a trajectory caused by uncertainty such as wind. A landing must keep at least 99% of the possible dispersions inside the landing zone. Read more: How Rockets Land Themselves
E
- Entanglement
- A link between qubits whose results are correlated: measure one and the result of the other is fixed. Read more: What Quantum Computers Can and Cannot Do
F
- Fine-tuning
- Further training of a model that is already trained, on a smaller and more specific set of examples, to change how it behaves. Read more: How a Language Model Is Trained
- Flag (feature switch)
- A setting that turns a piece of code on or off. At Knight Capital a reused flag switched on an old, forgotten function on one server. Read more: Knight Capital: 45 Minutes, $460 Million
H
- Hallucination
- A fluent answer that the facts do not support, such as an invented citation or event. Read more: Retrieval-Augmented Generation, Explained Plainly
- Hardware interlock
- A safety check built from hardware that stops a machine whatever the software says. The Therac-25 relied on software where earlier models kept these. Read more: Therac-25: When Software Was the Only Safety Net
- Humanoid robot
- A robot with a human-like body plan, usually a torso, two arms and two legs, built to work in spaces designed for people. Read more: How Humanoid Robots Actually Work
- Hyperscale data centre
- A very large facility run by a company that provides internet services and platforms at massive scale. Read more: Data Centres: Where the Power and Water Go
I
- Industrial robot
- Under the international standard ISO 8373:2021, an automatically controlled, reprogrammable, multipurpose manipulator, programmable in three or more axes, for use in automation. Medical robots are not counted. Read more: Industrial Robots by the Numbers
- Inertial reference system
- Sensors that tell a rocket how it is oriented and moving. Ariane 5 Flight 501 had two, and both failed from the same software fault. Read more: Ariane 5 Flight 501: Thirty-Seven Seconds
L
- Landing ellipse
- The region in which a landing vehicle is 99% likely to touch down. For Mars it has been measured in kilometres. Read more: How Rockets Land Themselves
- Logical qubit
- One more reliable qubit made by grouping many noisy physical qubits with error correction. It works only if the physical error rate is below a threshold. Read more: What Quantum Computers Can and Cannot Do
O
- Open source
- Software released under a licence that meets the Open Source Definition of the Open Source Initiative, which among other things bans discrimination against persons, groups or fields of endeavour. Read more: Open Source vs Open Weights in AI
- Open weights
- A model whose final trained weights are published so others can run and adapt it. This is not the same as open source: the training data and code, and the licence terms, may stay closed. Read more: Open Source vs Open Weights in AI
- Overflow
- What happens when a value is too large for the number of bits set aside to store it. In Ariane 5 a 64-bit floating-point number did not fit in a 16-bit signed integer. Read more: Ariane 5 Flight 501: Thirty-Seven Seconds
P
- Parameter
- A number inside a model that is adjusted during training. What a model has learned is stored in its parameters. Read more: How a Language Model Is Trained
- Parametric and non-parametric memory
- Parametric memory is what a model stores in its own parameters. Non-parametric memory is an external store it can consult, such as a document index. Read more: Retrieval-Augmented Generation, Explained Plainly
- Pretraining
- The first stage of training a language model: predicting the next token over huge amounts of text, adjusting the parameters whenever the guess is wrong. Read more: How a Language Model Is Trained
- PUE (power usage effectiveness)
- The total energy a data centre uses divided by the energy that reaches its computing equipment. Closer to 1 is better. Read more: Data Centres: Where the Power and Water Go
Q
- Qubit
- The basic unit of quantum computing. Unlike a bit, it can be in a superposition of 0 and 1 until it is measured. Read more: What Quantum Computers Can and Cannot Do
R
- Race condition
- A fault where the result depends on the timing of two tasks that share data. In the Therac-25, an edit made during the machine's setup was not noticed by the software. Read more: Therac-25: When Software Was the Only Safety Net
- Radiation-hardened processor
- A chip designed to keep working, or to detect and correct memory corruption, under the radiation found in space. Perseverance on Mars uses one, the RAD750. Read more: Why Space Computers Are So Old
- RAG (retrieval-augmented generation)
- A way to let a language model look up passages in an external collection of documents before it answers, so its knowledge can be updated by changing the collection instead of retraining. Read more: Retrieval-Augmented Generation, Explained Plainly
- Reward model
- A model trained to predict which of several outputs people would prefer. In RLHF it stands in for the human raters during reinforcement learning. Read more: What RLHF Does and Does Not Do
- RLHF (reinforcement learning from human feedback)
- Fine-tuning a model on human judgments: people rank its answers, a reward model learns their preferences, and the model is then trained to produce answers the reward model rates highly. Read more: What RLHF Does and Does Not Do
S
- STRIPS
- A planning method from the Shakey project, by Richard Fikes and Nils Nilsson. It describes each action by its preconditions and by what it adds to and deletes from a model of the world, and its descendants are still used in planners. Read more: Shakey, the Robot That Gave Us A*
- Superposition
- A quantum state that combines possibilities. NIST cautions that it is not "trying every answer at once" when the result is read out. Read more: What Quantum Computers Can and Cannot Do
T
- Terawatt-hour (TWh)
- A unit of energy equal to a billion kilowatt-hours. US data centres used an estimated 176 TWh in 2023. Read more: Data Centres: Where the Power and Water Go
- Token
- The small chunk of text a language model reads and writes. In English it averages about four characters, or three quarters of a word, so 100 tokens are roughly 75 words. Read more: How a Language Model Is Trained
W
- Workflow
- A system where models and tools follow code paths that the developer fixed in advance. More predictable than an agent, and often cheaper. Read more: What Is an AI Agent, Really?