Letter T

A hypothetical future point when technological change becomes extremely fast and difficult to predict.

The idea is often linked to AI improving AI, creating a rapid feedback loop. It is a speculation about the future, not an observed event or a date that experts agree will arrive.

#

A setting that influences how predictable or varied generated output is.

Lower temperature usually favours safer, more likely choices; higher temperature allows less likely choices and more variety. The exact behaviour differs between models and services.

#

A specialised Google-designed chip for accelerating machine-learning calculations.

A TPU is built to perform the large blocks of repeated mathematics common in neural networks efficiently. It is one type of AI accelerator, alongside GPUs and other specialised chips, and is better suited to some workloads than others.

#

An open-source software platform used to build, train and run machine-learning models.

TensorFlow is a set of programming tools rather than an AI model itself. It gives developers reusable building blocks for creating models and moving them from experiments into websites, servers, phones and other devices.

#

Examples kept untouched until the end to estimate how a finished model handles new cases.

Training data is homework and validation data is a practice exam. Test data is the final exam. If developers repeatedly look at it while changing the model, it stops being an honest test and the reported score may be too optimistic.

#

Using a written description to generate another kind of content, such as an image, video, audio, code or 3D object.

The X stands for the desired output. The model translates patterns in language into patterns in another medium, rather like giving a creative brief to a specialist. It does not retrieve a guaranteed perfect match: the result depends on the model, prompt, training data and generation settings and usually needs checking or revision.

#

A small chunk of information that a model reads or writes.

For text, a token may be a word, part of a word, punctuation or even spacing. Models process tokens rather than sentences exactly as people see them. Usage limits and prices are often measured in tokens.

#

Workplace slang for deliberately using very large amounts of AI tokens.

The phrase became associated with encouraging heavy AI use, sometimes through token-use leaderboards. Supporters see experimentation and adoption; critics point out that spending more tokens is not the same as producing better work. It is a trend, not a technical method or reliable productivity measure.

#

Letting an AI request outside capabilities such as search, calculation or file access.

A model alone mainly transforms inputs into outputs. Tools let the surrounding application fetch live facts or perform actions. Permissions and validation are essential because tool calls can have real effects.

#

A setting that limits token choices to a group with enough combined probability.

Also called nucleus sampling, top-p keeps the likely set flexible: a confident next step may have few choices, while an uncertain one may have many. It is another way to control variety.

#

Abusive, hateful, threatening or otherwise seriously harmful language produced or detected by a system.

Toxicity checks try to reduce harmful content, but language depends on context. A filter can miss disguised abuse or wrongly block people discussing abuse, reclaiming a slur or quoting material for legitimate reasons.

#

The process of adjusting a model so it learns patterns from data.

During training, the model makes predictions, measures error and changes its parameters to improve. This may happen over many examples and requires much more compute than a single use of the finished model.

#

The examples a model studies while learning.

Training data is the model's practice material. Its coverage and quality shape what the model learns, rather like a student whose understanding depends on the books, exercises and corrections they receive. Missing or skewed examples can lead to weak or unfair results.

#

A mismatch between how data is prepared during training and how it appears when the model is used.

It is like practising with measurements in centimetres but receiving inches in the real exam. Even a well-trained model can fail if the live system calculates features differently, omits information or presents a changed data format.

#

Reusing knowledge learned for one task to help with another related task.

A person who can ride a bicycle has a head start learning a motorbike. Similarly, a model that already recognises general image features can be adapted to identify a particular kind of object with less new data.

#

A neural-network design that powers most modern language models.

Transformers use attention to process relationships across a sequence and can be trained efficiently at large scale. They are used for language, images, audio and other kinds of data.

#

Being open about where AI is used, how it works and what its limits and effects are.

Transparency can include telling people they are dealing with AI, describing the data and testing used, naming who is responsible and explaining how a decision can be challenged. It is broader than explaining one prediction and should reveal useful information, not merely produce a long technical document.

#

A simple imaginary machine used to describe what computation can do in principle.

Proposed by Alan Turing, it reads and writes symbols on an unlimited strip according to precise rules. It is not a particular physical computer; it is a mathematical idea used to reason about algorithms and the limits of computing.

#

A proposed test of whether a machine can converse in a way that seems human.

Alan Turing suggested an imitation game in which a judge communicates by text and tries to distinguish a machine from a person. Passing such a test would show convincing conversation, not necessarily consciousness, truthfulness or general intelligence.

#