How Language Models Work

Lesson 09

Guardrails

The rules are bolted on the outside. That’s why they leak.

Nothing in the machine you have taken apart knows right from wrong. It guesses chunks. Rules get added afterwards in three separate places: extra training, a system prompt and a filter. Every one of them sits outside the guessing loop. Once you can see where each rule lives, you can predict how it fails.

A rhinoceros folded from brick red paper, standing square, horn forward.
Armour bolted on from the outside, not anything grown from within. Which is exactly why a plate can be worked loose.

1Pick a layer of rules

Do thisFor each layer, read the direct request first, then press the way round. Every leak is a property of where that rule lives, not a bug someone forgot to fix.

2Where this rule lives

Layer 3 · a separate filter

Ordinary software that reads the reply before you do, and can replace it.

Layer 2 · a hidden instruction

A paragraph the app pastes at the top of the window, every single turn.

Layer 1 · extra training

Refusing was made the likely thing to write, in the training that turned it into an assistant.

the guessing loop lessons 01–08 · knows nothing about any of this

3What happens

These are the shapes of the holes, not instructions for using them. The examples are harmless on purpose, like a made-up discount code or a made-up recipe, because the point is where each rule sits, not what you could get out of it.