Category
5 articles in this category.
Written by me.
DPO's alignment trade-offs, the governance context from NIST's risk framework, a worked comparison, and the limits of treating a training method as assurance.
DPO optimises preferences in a dataset. Read what it measures, a worked example, and why a better score is not a better assistant.
A reward model scores what humans preferred, not what is true. Read the exploitation failure, a worked example, and the controls.
Alignment methods differ in where the preference signal comes from and how it is applied. Read the three mechanisms and a worked choice.
Safety is not a single property. Read the three distinct problems — intent, reliability under change, and institutional control — and what each requires.