The model and the app around it

When you use an AI app, two layers are actually at work. The inner layer is the model, a large language model that takes text and produces text. The outer layer is the app: the chat screen, the upload button, the settings, and the tools attached around the model.

The model by itself cannot open the internet, cannot read files on your computer and does not know what time it is. All it has is the knowledge from its training, with a cut-off date (see the lesson Training and the knowledge cutoff), and the text sent to it right now. Every other ability is added by the app:

• Web search: the app searches, then pastes the results into the text the model reads.
• Reading files and images: the app turns files into text, or sends images to a multimodal model, one that can take images as well as text.
• Running code or calculating: the app runs code the model writes in a separate place, then returns the result.
• Connectors: the app links to your e-mail, calendar or file storage.

That is why the same model can feel very different in two apps. In one it can look up yesterday's news; in another it only guesses from its training memory. The first question whenever you use a new tool: which abilities does this app really provide, and which are switched on?

The most misunderstood thing is memory. The model remembers nothing from one chat to the next. Every time you send a message, the app sends the conversation, from the first message, to the model again; when it gets too long, older parts are cut or summarised. The model reads it from the start in its context window, then answers. Once the chat is closed, nothing remains in the model.

If the app seems to "remember" that you work in finance, that is because the app keeps memory notes and quietly pastes them at the start of each new chat. The settings are covered in the lesson on data settings and memory.

From this comes the list of things you must always give yourself: the goal and reader of your work, your organisation's rules and terms, the data or documents in question, and facts that changed after the model's knowledge cut-off, such as prices, new rules or who holds an office now. The model cannot guess these correctly. If you do not supply them, it fills the gap with something that sounds reasonable.

A quick way to tell which tools are active: look at the icons or buttons around the input box (usually there are options for search, file upload or a list of connectors), and watch the answer. An app that really searched usually shows clickable sources and mentions dates. An app that read a file can quote exact sentences from it. If the answer shows none of these signs, assume it came from the model's training memory. For answers that use sources, how to check them is in the lesson on research with sources. This small habit stops you from thinking an answer was checked against the outside world when the model only wrote something that sounds right.