AI from the beginning

Five things that are enough to begin

If you read only this section, you will know more than most of the people talking about AI.

1. 01

   AI is not programmed, it is trained

   An ordinary program does what somebody wrote it to do: if this, then that. AI is built the other way round. It is shown an enormous number of examples and works out the regularities itself. Nobody writes a rule for what a cat looks like. The machine looks at a million pictures of cats and infers it.

2. 02

   It guesses what comes next

   When you write to an AI, it does not look the answer up in a database. It produces the answer a word at a time, choosing at every point the most likely continuation given what it has read. The whole conversation comes out of that one trick, repeated on a machine trained billions of times over.

3. 03

   Likely is not the same as true

   From this follows AI's best-known fault. When the machine does not know something, it does not say so; it produces a continuation that sounds likely. The result can be entirely invented and still fluent and convincing. This is called hallucination, and it is a property rather than a bug: the machine does not distinguish a fact from a plausible turn of phrase.

4. 04

   It does not understand and it does not remember

   There is no place inside the machine where understanding happens. There are numbers that have been adjusted until the output is useful. It does not remember you between conversations either, unless somebody has built it a memory on purpose. Every conversation starts from an empty table.

5. 05

   Using it costs money every time

   AI is not a program you buy once. Every question and every answer consumes computation in a data centre somewhere, and somebody pays for it. That is why the price does not depend on how many users you have but on how much they use it.

Three rules that apply to any use

• Check whatever you are basing a decision on

  Names, figures, dates, legal provisions and sources. Those are exactly the things the machine invents most fluently.

• Do not put anything in that you would not send to an outsider

  Personal data, customer records and confidential material end up on somebody else's company's servers.

• Ask again in different words

  The machine does not take offence. If an answer feels vague, say more precisely what you want, who it is for and what shape it should take.

Language model

An AI trained on text that produces and handles language. The engine behind services like ChatGPT.

Prompt

What you write to the machine. The instruction and the question in one. A more precise prompt gets a more usable answer.

Hallucination

A confidently stated claim that is not true. Not lying, but a consequence of how the machine works.

Token

Text chopped into pieces. The machine reads and writes them, and they are what you are billed for.

Agent

A language model given tools and permission to try again. It does not merely answer; it acts.

Where to go from here

The order is deliberate. Try it first, read afterwards.