The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / S
15 terms, explained without the techno-murk.
The work of reducing the ways AI systems can cause harm.
AI safety includes reliability, misuse prevention, security, fairness, privacy, oversight and the effects of powerful future systems. Different risks need different practical safeguards.
A high-stakes system is tested for errors, protected from attack and monitored after release.
Choosing an output from the possibilities a generative model considers likely.
Rather than always selecting the single most likely next token, a system can sample among several. Settings such as temperature and top-p control how broad that choice is.
Sampling helps the same creative-writing prompt produce varied suggestions.
A technique that lets each part of a sequence weigh the relevance of the other parts.
It is like every word in a sentence looking around at all the other words and asking, 'Which of you helps explain me?' This helps a transformer connect references and ideas even when they are far apart.
In 'The animal was tired, so it stopped,' self-attention helps connect 'it' with 'animal.'
Learning from data by creating the practice question and answer from the data itself.
Instead of paying people to label every example, the system hides or predicts part of what is already there. It is like covering a word in a sentence and using the original word as the answer key.
A language model learns by predicting missing or next tokens in large collections of text.
Searching by meaning and intent rather than only matching the exact words typed.
Traditional keyword search looks for the same words. Semantic search tries to recognise that different phrases can express the same idea. It is like asking a librarian for books about 'keeping accounts safe' and also receiving a useful guide titled 'password security.'
A search for 'holiday allowance' finds a policy headed 'annual leave entitlement.'
Training with a small amount of labelled data and a larger amount of unlabelled data.
Imagine a teacher labels a few examples, then students use those clues while examining a much bigger unsorted pile. It can reduce expensive labelling work when unlabelled data is plentiful.
A medical model learns from a few scans labelled by doctors plus many scans without labels.
Estimating whether language expresses a positive, negative or neutral attitude.
It is like asking a computer to read the mood of a comment. The result is an estimate, not a mind-reading device. Sarcasm, jokes, cultural differences and mixed feelings can easily confuse it.
A company groups thousands of product reviews by broadly positive, negative or neutral sentiment.
A language model designed to use fewer resources than a large one.
An SLM can be faster, cheaper and easier to run privately or on a device. It may be the better choice for a focused task even if it has less broad ability than a large model.
A small model handles offline text suggestions on a phone.
Turning spoken language into written text or commands.
The system analyses sound patterns and estimates which words were spoken. Accents, noise, overlapping voices and unfamiliar names can make the task harder.
A phone transcribes a voice message into text.
The people or groups who build, use, oversee or may be affected by a system.
Stakeholders include more than customers and developers. They may include workers, communities, people represented in the data, regulators, suppliers and those who never chose to use the system but are affected by its decisions. Good governance involves them before harm occurs, not only after launch.
For an AI hiring tool, stakeholders include applicants, recruiters, current workers, equality specialists and regulators.
Creating an image that keeps the subject of one picture but borrows the visual style of another.
Imagine repainting a holiday photograph using the colours, textures and brushwork of a reference painting while keeping the buildings and people in place. The system separates features associated with content from features associated with style, though the boundary between them is not always clean.
A photograph of a city street is re-created with the appearance of a watercolour illustration.
Learning from examples that come with the correct answers.
It is like practising with an answer key. The model makes a guess, compares it with the supplied label and adjusts so future guesses improve.
A model learns from photographs labelled 'cat' or 'dog' before classifying new photographs.
AI built from explicit symbols, facts and logical rules.
Instead of learning everything from examples, symbolic AI represents knowledge in a form people can often read: facts, categories and if-then rules. Modern systems sometimes combine this approach with neural networks.
A rule states that if all humans are mortal and Ada is human, then Ada is mortal.
Artificially created data used in place of, or alongside, real-world data.
Synthetic data can fill gaps, create rare examples or reduce some privacy concerns. It can also repeat errors and biases from the system that generated it.
A driving simulator creates thousands of unusual road situations for training.
Higher-priority instructions that shape how an AI assistant should behave.
A system prompt may set the assistant's role, boundaries, tone and tool rules. It is usually supplied by the product maker rather than typed by the end user.
A travel assistant's system prompt tells it to ask before making any booking.