I think this is exactly where alignment gets difficult. It’s not just about preventing the model from jumping guardrails. It’s about setting the right constraints so it consistently chooses to operate within them.
But then comes the harder question: whose rules?
Easy examples create false confidence because most of us agree on the answer. The real alignment problem lives in edge cases where values conflict, context matters, and reasonable people disagree.
That’s where I think alignment stops being purely a model problem and starts becoming a governance problem.
Show more
