The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / E
12 terms, explained without the techno-murk.
A list of numbers that represents the meaning or features of something.
Embeddings place similar items near each other in a mathematical space. They help computers search by meaning rather than relying only on exact matching words.
An embedding search can connect 'How do I reset my password?' with a guide titled 'Recover account access.'
AI that learns or acts through a body in a physical or simulated world.
Rather than only reading text, embodied AI connects sensing, movement and consequences. A robot learns that the world pushes back: objects have weight, doors have handles and actions can fail.
A household robot uses cameras and touch sensors to learn how to pick up unfamiliar objects.
An ability or pattern that seems to appear as an AI system grows or becomes more complex.
Some capabilities look sudden at a certain scale, although the appearance can depend on how tests are scored. The term does not mean the system became conscious or developed the ability by magic.
A larger model may begin handling a task that smaller versions performed poorly on.
An encoder turns input into a useful internal representation; a decoder turns a representation into output.
Think of the encoder as packing the important meaning into a suitcase and the decoder as unpacking it into the form needed at the destination. Some models use both parts, while others use mainly an encoder or mainly a decoder.
A translation system encodes an English sentence's meaning and decodes it as a French sentence.
Turning readable information into a protected form that requires a key to unlock.
Encryption helps keep data private while it is stored or sent. It is a security technique rather than a type of AI, but it matters whenever AI systems handle confidential information.
A messaging app encrypts a conversation so an unauthorised person cannot read it in transit.
Training one system to learn the whole journey from raw input to final answer.
Instead of people designing each intermediate step, the model learns the connections for itself. It is like teaching someone to turn raw ingredients directly into a finished meal rather than handing them separately prepared pieces. This can be powerful, but the middle of the process may be harder to inspect.
A speech system learns to turn an audio recording directly into text without a person specifying every sound feature.
Combining several models so their joint answer is better than relying on one alone.
It is like asking a panel instead of one judge. If the models make different mistakes, voting or combining their estimates can produce a steadier result.
A fraud system combines several detectors and flags a payment when their overall evidence is strong.
One complete pass through the training dataset.
Think of an epoch as working through an entire practice workbook once. Models usually make many passes, improving each time, though too many can lead to memorising the workbook rather than learning the subject.
Ten epochs means the training process has worked through every example ten times.
A framework for judging how developed an organisation's responsible-AI practices are.
It works like a progress map. An organisation may move from informal good intentions toward documented checks, clear ownership, regular testing and independent oversight. Reaching a high level on a framework does not prove that every AI system is ethical; evidence and continuing review still matter.
A company assesses whether it records model risks, tests for bias and gives customers a way to challenge automated decisions.
Thinking carefully about how AI should be designed, used and governed.
AI ethics asks practical questions about fairness, privacy, responsibility, human choice, environmental cost and who benefits or bears harm. It is not a single checklist with universally agreed answers.
A council considers whether a face-recognition system is accurate, necessary, fair and open to challenge.
Tests used to measure how well and how safely an AI system behaves.
Short for evaluations, evals can be automated test sets, expert reviews or realistic trials. Good evals test the actual job and its possible failures rather than relying on one headline score.
A medical-summary eval checks whether important facts are retained and whether invented details appear.
Software that uses a collection of specialist rules to make recommendations in a narrow field.
An expert system is closer to a detailed decision handbook than a modern generative model. Human experts supply rules such as 'if these conditions are true, consider this conclusion.'
An early medical expert system asks about symptoms and follows rules to suggest possible diagnoses.