AI Anatomy Class · Part 8 · The birth of ChatGPT

79. Making AI safe

Q. Which requests should an AI refuse? And what if it refuses too much?

1. Huge numbers of people use ChatGPT every day. Mixed in are requests from people who want to use it to harm others.
2. Making a model act in line with what people intend and value is called alignment. Chapters 73 and 76 were its tools.
3. It's taught to refuse dangerous requests, like secretly getting a friend's password, and to say what it can do instead.
4. But being too scared is a problem too. Refusing to explain chopping onions just because of the word 'knife' is useless.
5. So it needs balance. Too strict, and it refuses fine requests. Too loose, and it helps with dangerous ones.
6. A good refusal is short and polite. No lecturing: give a brief reason and point to another way it can help.
7. Part 8 recap! A base model plus model answers, rankings and a grader, praise, and a chat format became ChatGPT.
8. Now it's time to watch an answer being made. In Part 9 we look closely at how the next piece is chosen.

Chapter summary

• Alignment: making a model act in line with what people intend and value.
• Refuse dangerous requests, but don't refuse fine ones just because of a word.
• A good refusal is short and polite, and offers another way to help.