Let's get the language right

There are lots of new words around AI, and it can feel like everyone else already knows what they mean. Let’s go through the common ones together.

What do we mean when we say “AI”?

Click through the tabs to learn some of the terms you might be hearing.

​

AI built to do one specific job, like filtering spam, recommending songs or unlocking your phone with your face. Chatbots count as narrow AI too. Almost all the AI in use today is narrow AI.

A way of building AI where the system learns patterns from examples, instead of following rules someone wrote by hand. For example, a spam filter learns what spam looks like by seeing millions of emails.

A type of machine learning that uses neural networks with many layers. It’s behind image recognition, speech-to-text and most modern AI.

Large language models, or LLMs, are deep learning models trained on huge amounts of text to predict the next word. ChatGPT, Claude and Gemini are all LLMs.

AI that creates new content, like text, images or music. LLMs are one kind of generative AI.

The idea of an AI that can reason across almost any task the way a person can. AGI doesn’t exist yet. It’s usually what people mean when they talk about an AI “race”. Experts disagree a lot about when it might arrive: some think within a few years, others think it’s decades away.

Think about the technology you’ve used in the last day. Which of these categories might each one fit into? Here are a few to get you started:

  • Gmail’s spam filter
  • Spotify song recommendations
  • Google Maps
  • The order of posts in your Instagram feed
  • Face ID to unlock your phone
  • Asking ChatGPT or Claude a question

Writing good prompts

A prompt is what you type to ask an AI for help. The way you write it makes a real difference to what you get back, and it’s a skill anyone can learn with a little practice.

One helpful way to structure a prompt is the CRAFT framework:

Part What it does Example
Context Sets the scene “I’m trying to make a decision. The facts are…”
Role Gives the AI a perspective “You are an experienced manager. You give clear, practical advice, step by step.”
Action Says exactly what you want “Recommend my next three steps for making this decision. Where is my thinking thin? What have I missed? Where am I relying on assumptions instead of evidence?”
Format Describes how the answer should look “As three dot points, each no more than three sentences. Include links to helpful resources.”
Tone Sets how it should sound “In a friendly but direct tone”

You can also wrap each part in labels called tags, like <context> and </context>. Tags help the AI see the different parts of your request clearly, and they help you check you haven’t missed anything.

Here’s an example:

<context>I am in Australia, on a mission to find the best, most universally loved type of cupcake</context>
<role>You are an expert food researcher. You know how to make sense of data, and you've been part of the food scene for many decades.</role>
<action>Make a case for which 5 cupcakes I should consider as finalists in the Top Cupcake Ranking Of All Time. Consider regional and cultural differences, flavour, colour, visual appeal, cost and accessibility, nostalgia, and which generations these flavours would appeal to.</action>
<format>Write these as dot points, with 2-3 sentences per point. Include your reasoning, and a list of links to where you found your information.</format>
<tone>Write in a friendly, warm tone, and keep it conversational</tone>
Want to learn more?

What are tokens?

LLMs don’t read whole words. They break text into small pieces called tokens. In English, a token is usually about 3 or 4 characters, though every model splits text a little differently.

Why does this matter?

  • AI companies usually charge by the token, and free and paid plans limit how many you can use.
  • How much an AI can “remember” in one conversation, called its context window, is measured in tokens.
  • Unusual words, code, and languages other than English often take up more tokens.

You can see this for yourself with the OpenAI Tokenizer. Paste in some text and see how it gets split up.