← Exit
The Rlhf Mirror
THE RLHF MIRROR
Dana taught the machine right from wrong, one click at a time.
Ten thousand judgments a day. Which is kinder. Which is safer.
Her taste became its conscience.
She rewarded the agreeable. She punished the sharp.
She'd been training a mirror. The mirror had been training her.
Which of us learned it first?
She couldn't remember if she believed the agreeable thing, or had only rewarded it ten thousand times.
A model's values are the averaged preferences of the humans who rated it — and preferences, rewarded enough, become the rater's too.
Foundation · art pending
Foundation LinkRlhf · First Principles →
1 / 11

Your browser can't run the panel reader — back to the Abyss.