The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / F
10 terms, explained without the techno-murk.
A false positive is a false alarm; a false negative is a real case the system missed.
A burglar alarm sounding when nobody broke in is a false positive. Failing to sound during a real break-in is a false negative. Which error matters more depends on the situation and should shape the system's threshold.
A spam filter may wrongly block a genuine email or allow an actual scam message through.
A piece of information a model uses when making a prediction.
Features are the clues given to the model. For a house-price model, floor area and location might be useful features; the colour of the front door probably is not.
Age, account history and recent activity may be features in a fraud model.
Choosing or creating useful clues from raw data for a model to learn from.
Raw information is not always in the most helpful form. Feature engineering is like preparing ingredients before cooking: dates might become 'days since last purchase', and several measurements might become one meaningful ratio. Poorly chosen features can hide the real pattern or introduce unfairness.
A fraud model uses the distance from a customer's usual location as a feature rather than the raw address alone.
Training across many devices or organisations without gathering all their raw data in one place.
It is like sending the lesson to several classrooms and collecting their suggested improvements rather than collecting every student's private workbook. It can reduce data sharing, though it does not remove every privacy or security risk.
Phones help improve a typing model while keeping the users' actual messages on their devices.
Helping a model perform a task using only a small number of examples.
In a prompt, a few worked examples can demonstrate the desired pattern, rather like showing someone two completed forms before asking them to fill in a third. More broadly, few-shot learning means adapting from very limited task-specific data.
A prompt shows three reviews with sentiment labels before asking the model to label a new review.
Extra training that adapts an existing model for a particular task or style.
Fine-tuning starts with a model that already knows broad patterns, then teaches it using a smaller, focused dataset. It changes the model itself, unlike simply adding information to a prompt.
A support team fine-tunes a model on approved examples of its preferred reply style.
Passing an input forward through a neural network to produce an output.
Imagine information moving along an assembly line. Each layer transforms what it receives and hands the result to the next layer until the network produces its answer. During training, that answer is checked, and backpropagation then traces any error in the opposite direction.
An image moves through a network's layers until the final layer produces probabilities for 'cat' and 'dog.'
A broadly trained model that can be adapted for many different uses.
Foundation models are trained at large scale on varied data. Apps can then steer them with prompts, tools, retrieval or fine-tuning instead of building a separate model from scratch for every task.
One foundation model might power writing, classification, search and customer-support tools.
A model near the leading edge of current AI capability.
The phrase usually refers to the most capable general-purpose models available at a given time. There is no permanent list or single agreed threshold, and today's frontier will not remain the frontier for long.
Governments may apply extra testing requirements to frontier models because of their broad capabilities.
A way for a model to request a specific software action using structured information.
Instead of pretending it completed an action in prose, the model produces a request that ordinary software can validate and run. The result can then be returned to the model.
A model requests `check_weather` with a city and date, then explains the real forecast it receives.