The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / A
35 terms, explained without the techno-murk.
Comparing two versions to see which performs better with real users or data.
One group receives version A and another receives version B, while everything else is kept as similar as possible. It is like offering two shop-window designs on alternate streets and measuring which attracts more visitors, rather than relying on opinion alone.
A company tests whether an AI-written subject line gets more email opens than its existing one.
A test that removes one part of an AI system to see how much that part matters.
It is like taking one ingredient out of a recipe and tasting the result. If performance drops sharply, that ingredient was important; if little changes, it may not have been doing much.
Researchers switch off a model's memory feature and rerun the same tests to measure the difference.
Designing technology so people with different abilities can use it successfully.
Accessibility should be considered from the beginning, not added as a last-minute repair. It can involve clear language, keyboard control, captions, screen-reader support, good colour contrast and flexible ways to give input or receive information. AI can improve access, but it can also create new barriers when it is inaccurate or excludes disabled people from testing.
An AI meeting tool provides accurate live captions and lets users correct names the system misheard.
The share of a model's predictions that were correct overall.
Accuracy is simply correct answers divided by all answers. It can be misleading when one outcome is rare: a system that always predicts 'no flood' may be almost perfectly accurate while failing to warn about every actual flood.
If a model gets 80 out of 100 classifications right, its accuracy is 80 percent.
A small mathematical rule that decides how strongly a unit in a neural network should respond.
Think of it as a dimmer switch rather than a simple on-off button. It transforms an incoming signal before passing it onward, helping the network learn complicated patterns.
An activation function may pass a strong useful signal forward while reducing a weak one.
Training where the model asks people to label the examples it would learn most from.
Instead of labelling everything, the system points to the cases it finds confusing or informative. It is like a student choosing to ask the teacher about the questions they are least sure of, making limited teaching time more useful.
A medical-image model sends its most uncertain scans to specialists for labelling.
Technology designed to recognise, respond to or imitate human emotion.
It may analyse words, facial expressions, voice or body signals. It is making a calculated guess about emotion, not reading a person's mind, and those guesses can be culturally biased or simply wrong.
A call-centre system estimates whether a caller sounds frustrated and alerts a supervisor.
Coordinating several AI agents or model calls so they work together on a larger task.
It is similar to a project manager dividing work among specialists, passing results between them and deciding what happens next. More agents do not automatically mean a better result; poor coordination can multiply cost and mistakes.
One agent gathers evidence, another drafts a report and an orchestrator sends weak sections back for revision.
The repeating cycle an AI agent follows while working toward a goal.
A typical loop is: observe what is happening, decide what to do, act, inspect the result and repeat. It resembles using a satnav: check the current position, choose the next turn, move, then recalculate from the new position.
An agent searches for a train, checks the results, changes its search and continues until it finds a suitable journey.
A computer chip designed to perform the heavy calculations used by AI quickly.
A normal processor is a capable general worker. An AI accelerator is more like a large team doing similar sums at the same time. GPUs and TPUs are common examples.
A data centre uses thousands of AI accelerators to train a large model.
An AI system that can work through steps and take actions toward a goal.
A normal chatbot mainly replies. An agent can also plan, use tools, check results and decide what to do next. Its freedom can range from tightly supervised to fairly independent.
An agent might read new support emails, look up orders and draft replies for a person to approve.
The effort to make AI systems behave in line with human intentions and values.
Alignment asks whether an AI is doing what people actually meant, not merely following the most literal version of an instruction. It includes technical safety, social choices and difficult questions about whose values count.
A well-aligned assistant should refuse a harmful request even if the request is phrased cleverly.
Units used by some AI services to measure and limit how much of the service a person can use.
Credits are like prepaid tickets at a fair. A simple task may use one ticket, while generating a long video or using a powerful model may use several. Each provider defines credits differently, so they cannot usually be compared directly across services and are not the same thing as tokens.
An image service gives a user 100 monthly credits and charges more credits for high-resolution generations.
A facility full of computers and networking equipment built for demanding AI work.
It is the physical home of the compute behind many AI services. These sites need large amounts of electricity, cooling, high-speed connections and specialised chips.
Training a frontier model may involve many racks of connected AI accelerators inside a data centre.
The effort to make sure an AI system does not unfairly disadvantage particular people or groups.
Fairness is not always one simple mathematical rule. Treating everyone identically can still be unfair when groups face different circumstances or error rates. Teams must decide which harms matter, test outcomes across relevant groups and give people ways to question important decisions.
A hiring system is checked to see whether equally qualified applicants receive similar results across demographic groups.
The rules, responsibilities and checks used to control how AI is chosen, built and used.
Governance answers practical questions such as: Who is accountable? What uses are forbidden? How will risks be tested? Who can stop the system? Think of it as the combination of road rules, driving tests, maintenance schedules and named owners that makes powerful machinery manageable.
An organisation requires an impact review, a responsible owner and regular monitoring before an AI service can affect the public.
The stages an AI system passes through from first idea to final retirement.
The lifecycle commonly includes planning, gathering data, building or adapting a model, testing, deployment, monitoring and eventually shutting the system down. It is more like caring for a building than buying a toaster: performance, users and risks can change, so maintenance continues after launch.
A benefits assistant is designed, tested with users, monitored for harmful errors and later retired when its rules become outdated.
The possibility that an AI system will cause harm, together with how serious that harm could be.
Risk depends on both likelihood and impact. A small spelling error and a mistaken medical decision are not equivalent, even if they happen equally often. Good risk work asks who could be harmed, how the failure might happen and what safeguards or alternatives reduce it.
A hospital treats an unreliable diagnosis as a higher risk than an imperfect suggestion for rewording an email.
Low-quality AI-generated material produced quickly and in large quantities.
The phrase usually describes content made to fill space, attract clicks or overwhelm a platform rather than inform or delight anyone. The problem is not simply that AI was used; it is that quantity was valued above care, checking and originality.
A website publishes hundreds of repetitive, unchecked AI articles containing invented facts.
A set of steps a computer follows to solve a problem or make a decision.
A recipe is a useful comparison: it takes ingredients or inputs, follows instructions and produces a result. AI systems often use many algorithms together.
A route-finding algorithm compares possible roads and chooses a journey.
When a computer system repeatedly produces unfair or skewed results.
The problem may come from biased historical data, missing groups, unsuitable rules or the way the system is used. The computer does not need to intend unfairness for its decisions to disadvantage people.
A recruitment system trained on an unbalanced history may rank equally capable candidates differently.
Finding cases that look unusually different from the normal pattern.
It is like noticing one bright red sock in a drawer full of black ones. An unusual case may signal fraud, a fault or an important discovery, but it may also be an innocent exception that needs checking.
A bank flags a payment that is far larger and in a different country from a customer's usual spending.
Treating something non-human as if it has human feelings, intentions or understanding.
AI can use warm language, imitate emotion and say 'I', which makes it easy to imagine a person behind the screen. That conversational style does not prove the system feels, believes or understands things as a human does.
A chatbot says it is sorry, but this does not mean it experiences regret or sadness.
A doorway that lets one piece of software ask another piece of software to do something.
API stands for application programming interface. An AI API lets an app send instructions or data to a model and receive the model's answer in a predictable format.
A travel website can use an AI API to turn a customer's notes into a tidy itinerary.
A proposed kind of AI able to handle a very wide range of intellectual tasks.
There is no single agreed test for AGI. People use the term for a future system that could learn and reason across many fields rather than being built for one narrow job. Claims that AGI has arrived should be treated carefully.
Unlike a chess program, an AGI would supposedly adapt to unfamiliar work across science, writing, planning and more.
Computer systems made to do tasks that normally need human-like judgement.
AI is an umbrella term. It covers everything from software that spots fraud to tools that write, draw, translate or hold a conversation. It does not necessarily think or understand in the way a person does.
A photo app recognising faces and a chatbot answering a question are both uses of AI.
A hypothetical AI that would be far more capable than humans across most intellectual tasks.
ASI is an idea about a possible future system, not a technology known to exist today. Imagine an intelligence that could outperform the best human specialists in science, strategy, invention and many other fields at once. Experts disagree about whether it can be built, when it might happen and what the consequences would be.
A fictional ASI might solve research problems that entire teams of leading scientists could not solve.
A technique that helps a model weigh which parts of its input matter most right now.
Attention lets a language model connect words and ideas even when they are far apart in a passage. It is a central ingredient of transformers, though it is not the same as human attention.
In 'The trophy did not fit in the suitcase because it was too big,' attention helps connect 'it' with the trophy.
Using AI to support human judgement rather than replace it.
The computer handles jobs such as scanning large amounts of data or suggesting possibilities, while a person adds context, values and responsibility. Think of a calculator for judgement: it can extend what someone can do, but the person still decides what the result means and what action to take.
An AI highlights unusual areas in a medical scan, then a clinician reviews the evidence and makes the diagnosis.
A neural network that learns to squeeze information down and then rebuild it.
Imagine summarising a picture into a compact set of notes, then trying to redraw the picture from those notes. By learning what must be kept, an autoencoder can discover useful features, reduce noise or spot unusual cases.
An autoencoder learns the normal pattern of machine sounds and flags recordings it cannot reconstruct well.
Software that automates parts of choosing, building and tuning a machine-learning model.
It is a toolkit that tries several recipes and settings for you. It can save time, but people still need to define the right problem, check the data and judge whether the result is safe and useful.
An AutoML service compares several prediction methods and recommends the best performer on a test set.
The tendency to trust a computer's recommendation too readily, even when it is wrong.
People can mistake an automated answer for an objective one or stop checking because the system usually works. A human review step is not useful if the reviewer simply clicks approve without genuine attention or authority.
A clinician overlooks conflicting evidence because a diagnostic system displayed a confident suggestion.
A machine or software system that can make some decisions and act without constant human direction.
Autonomy comes in degrees. A system may handle a narrow routine by itself while still needing human goals, permissions and help with unusual cases. Think of cruise control compared with a driver: it can manage one part of the journey, but that does not make it independent in every situation.
A warehouse vehicle plans a route around obstacles and stops when its sensors detect a person.
A vehicle that can sense its surroundings and handle some or all driving tasks itself.
It combines cameras or other sensors with maps, planning and control software. 'Autonomous' covers several levels, from driver assistance to a vehicle that needs no human driver in specific conditions.
A self-driving taxi identifies lanes and pedestrians, plans a route and controls steering and braking.
A model that creates the next part of its output using the parts it has already produced.
It works rather like continuing a sentence one piece at a time. Each new token becomes part of the context for choosing the next token, so an early choice can shape everything that follows.
A language model writes a paragraph token by token, repeatedly using its existing text to predict what comes next.