How AI Works: Three Ideas – free cheat sheet (episode 25)

Why it matters

A chat AI can feel like magic. Underneath, it is three plain ideas, plus one catch. You can picture each one in a kitchen with one sentence: Crack two eggs into the ...

The three ideas, plus the catch

0 of 4 done

A library of textBefore you ever chat, an AI is trained on a huge amount of text. Training means practicing on it. The library is an analogy, not real books.Try: Picture a giant pile of recipes.
Guess the next wordIt plays one game: guess what comes next, add it, guess again. Really it guesses tokens, which are words or pieces of words.Try: Crack two eggs into the ___
The catch: practice is not enoughAn AI that only practiced guessing is not automatically good at answering questions or following instructions.
Add manners classFine-tuning is more training. In one well-known method, people compare answers, and the AI learns to prefer the ones people liked.Try: Which answer would you pick: rambling or clear?

Tap a step when you have done it. Your ticks stay on this device.

Rule of thumb

AI is a well-read guesser with good manners.

Sources: public glossaries from AI makers (pretraining, fine-tuning, human feedback, tokens); a free online machine-learning course on language models; Ouyang et al., 'Training language models to follow instructions with human feedback' (arXiv 2203.02155). The kitchen and robot scenes are our own illustration.

Episode 25: How AI works in three ideas

Why it matters

The three ideas, plus the catch

More like this

A new free sheet every week