How Language Models Work

Lesson 14

The assistant act

It couldn’t answer you at first. That part was trained in afterwards.

A model fresh out of training does not answer questions. It continues text, like autocomplete that has read most of the internet. Being a helpful assistant is a second, separate round of training added on top, called fine-tuning.

A dog folded from camel paper, sitting attentively with its ears pricked.
Straight out of training it is a wild thing that finishes sentences. Being a helpful assistant is a second round, added on top — and this is what that round produces.

1Pick a message to send it

Do thisFlip between the two buttons below. Same machine, same knowledge. The only difference is a round of training that came afterwards.

2The same model, before and after that training

3How that second round of training worked

Step 1 · copy the shape

Show it tens of thousands of example conversations, so “an answer” becomes the likely thing to write next.

Q: How do I boil an egg?
A: Place the egg in boiling water for 7–9 minutes…

Q: Summarize this paragraph.
A: Here's the key point in two sentences…

Q: Write a limerick about soup.
A: There once was a chowder so thick…

Step 2 · let people pick a winner

Give it two tries at the same question, let real people choose the better one, and train it toward the winners.

✓ people picked this
“I'd check the oven temperature first. Uneven browning usually means it runs hot.”
“As an AI language model, baking is a complex topic with many variables to consider…”