AI basics

There is a lot of mystery and marketing murmur around AI. The basics are not hard, though.

Here they are explained from the start, step by step and without buzzwords. The videos are schematic examples, not live queries.

How does an AI get a task done?

An everyday example: a walk. There are three variants.

  • It answers directly. Only text comes back.
  • It uses a tool. The AI makes a request, the app runs it and returns the result.
  • It completes a task. The AI uses the result to choose the next step.

The AI decides, the app executes, only within the permissions and limits of the app.

A walk shows how AI can reply directly, use a tool or work through a task one step at a time.

Schematic example. The animation does not run real queries.

Show transcript

How does AI actually do stuff? A walk, three ways. An illustrative example using example data.

1. Just replies. Question: “When should I go for a walk?” Answer: “Try a cooler part of the day.” A text answer with no tools used.

2. Uses a tool. Question: “Will it rain this afternoon?” The AI requests the afternoon forecast. The app checks and returns: “Dry from 3 to 5 pm.” Answer: “3 to 5 pm looks good.”

3. Works through a task. Task: “Find a dry hour when I’m free.” The AI uses the forecast and checks the calendar. 3 to 4 pm is busy, so it tries 4 to 5 pm. That hour is free. Answer: “4 to 5 pm works.”

The AI chooses. The app acts within its permissions and limits. A run can also ask, fail or stop.

Chat. Tools. Agents.

Three ways to use a language model.

Chat replies directly. With tools, the app runs an action. Agents use results to choose the next step.

Schematic example. The animation does not run real queries.

Show transcript

Chat. Tools. Agents. Three ways to use a language model.

Text only. Your prompt goes to the model. The model generates a response.

One tool call. The model generates a request for a tool, such as a calculator. The app runs the tool and returns the result. The model uses that result in its answer.

Agent loop. Example task: “Find a price and calculate the total with tax.” The model can request tools such as search, read page and calculate. After each result, it chooses what to do next.

The model requests. The app executes. The app controls permissions and limits. Runs can finish, ask for help, fail or hit a limit.