The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / M
10 terms, explained without the techno-murk.
A way of building software that learns patterns from examples.
Traditional programming gives a computer detailed rules. Machine learning gives it data and a learning method, allowing it to form a model that can make predictions on new cases.
A spam filter learns from messages labelled 'spam' and 'not spam.'
Principles and practices that give Māori a meaningful role and authority in decisions about Māori data.
Māori data includes information about Māori people, language, culture, resources and environments. Governance goes beyond ordinary privacy: it considers collective rights, tikanga, wellbeing, control and Māori participation in the institutions and decisions that shape how the data is stored and used.
An agency involves Māori representatives in deciding whether and how Māori health data may be used to develop an AI service.
Information saved so an AI application can use it in a later interaction or step.
A model's context is its current desk; memory is material placed in a filing cabinet for possible later retrieval. The application may store summaries, preferences or past events and bring selected pieces back into context. This is not necessarily human-like remembering, and stored information can be incomplete or outdated.
An assistant remembers a user's preferred writing tone and retrieves that preference during a future conversation.
A measurement used to judge one aspect of an AI system's performance.
Metrics are scoreboards, and every scoreboard rewards something particular. A model can look excellent under one metric and poor under another, so the chosen measurement should match the real-world goal and cost of mistakes.
Accuracy, response time and the rate of harmful answers are three different metrics.
Misinformation is false information shared without necessarily intending harm; disinformation is deliberately misleading.
The same false claim can move from one category to the other as it spreads: one person may invent it deliberately, while many others repeat it believing it is true. Generative AI can make misleading text, images or audio faster to create, but people and distribution systems determine how it is used and amplified.
A fabricated photograph is posted to deceive voters, then reshared by someone who mistakenly believes it is genuine.
A model design that activates only selected specialist parts for each input.
A routing system sends each token to a few 'expert' networks instead of using every part of the model every time. This can increase capacity without increasing the cost of every response by the same amount.
Different parts of a model may become more useful for code, translation or factual questions.
A learned mathematical system that turns inputs into predictions or outputs.
A model is the part that has learned patterns during training. A chatbot or image app is a larger product built around one or more models, with an interface, rules and other software.
A weather model takes measurements as input and predicts tomorrow's temperature.
An open standard for connecting AI applications to tools and sources of information.
MCP gives different systems a common way to offer files, databases and actions to an AI application. It is often compared to a standard plug shape: it reduces the need for a separate custom connection every time.
An AI assistant uses MCP connections to read approved documents and query a project tracker.
A system in which several agents interact, cooperate or compete.
Each agent may have its own role, information or goal. Working together can divide a complicated task, but it also creates extra coordination problems and more opportunities for errors to spread.
One agent researches a topic, another drafts a report and a third checks the draft against the sources.
Able to work with more than one kind of information, such as text, images and audio.
A multimodal model may accept several media types, produce several types or connect information across them. Its ability can differ greatly by mode.
You show a model a chart and ask a spoken question about what it means.