An AI system set up to pursue a goal through a series of steps, using tools like a browser or a code editor.
8 entries where it mattered
Glossary
The words the record uses, defined for a reader who does not build any of this. Entries link here on first use.
A
The research problem of making AI systems do what people intend, and nothing else.
2 entries where it mattered
A defined way for one program to use another over the internet; how most developers reach AI models.
8 entries where it mattered
B
A fixed set of test questions used to compare models. Useful, gamed often, and frequently misread.
8 entries where it mattered
C
How much text a model can consider at once, measured in tokens. What falls outside it is forgotten.
6 entries where it mattered
D
The technique behind most AI image generators: start with noise, remove it step by step until a picture emerges.
1 entry where it mattered
Training a small model to imitate a large one, so it answers nearly as well at a fraction of the running cost.
2 entries where it mattered
F
A second, smaller round of training that adapts a finished model to a particular job or behavior.
Not used in an entry yet
A large model trained once at great expense, then adapted by many others for many purposes.
2 entries where it mattered
H
A fluent, confident model output that is simply wrong, invented rather than recalled.
1 entry where it mattered
I
Running a trained model to get an answer, as opposed to training it in the first place.
1 entry where it mattered
L
A model trained on very large amounts of text to predict what word comes next, and used to write and answer.
8 entries where it mattered
M
A design where a model holds many specialist sub-models and consults only a few for each token.
5 entries where it mattered
A program whose behavior was learned from examples rather than written out as rules by a person.
8 entries where it mattered
A model that handles more than one kind of input or output: text plus images, audio, or video.
7 entries where it mattered
O
A model whose trained parameters are published, so anyone can download and run it on their own hardware.
8 entries where it mattered
P
One of the adjustable numbers inside a model. Modern language models have billions of them.
8 entries where it mattered
Q
Shrinking a model by storing its numbers with less precision, trading a little accuracy for much lower cost.
Not used in an entry yet
R
Deliberately attacking a model to find its failures before the public does.
1 entry where it mattered
Training by trial, error and reward rather than by copying labelled examples.
8 entries where it mattered
Giving a model a search step: it looks up documents first, then writes its answer using what it found.
Not used in an entry yet
Reinforcement learning from human feedback: tuning a chatbot with people’s rankings of its answers.
Not used in an entry yet
T
The unit a language model reads and writes in: a word, part of a word, or punctuation.
8 entries where it mattered
The process of showing a model examples and adjusting its parameters until its answers improve.
8 entries where it mattered
The network design, published by Google in 2017, that almost every large language model is built on.
6 entries where it mattered