Hi, welcome to the EA Forum. It’s nice to see philosophical ideas that don’t come from the dominant tradition here.
Your argument rests on the premise that everyone (human) has liangzhi but large models don’t.
I’m skeptical of that, because the innate sense of right/wrong can be culture dependent, and there are people with neurological and psychological conditions that don’t have that same experience.
Thank you for reading carefully — these are genuinely interesting questions.
On cultural dependency:Liangzhi is not a specific set of moral beliefs — those do indeed vary across cultures. Liangzhi is the capacity for judgment that exists before any moral beliefs. Different cultures disagree on many things, but seeing an innocent child suffering and feeling an instinctive resistance to it — that is cross-cultural, and no one has to teach it. Liangzhi is this capacity itself, not any particular moral code.
On neurological and psychological conditions: This is a genuine edge case worth taking seriously. In my framework, these individuals don’t lack liangzhi — rather, their liangzhi is obscured by something very thick. Disease is one form of obscuring, just as siyu (private desires) is another form described in the post. I’ll admit this edge case deserves more careful thought. Modern science offers many explanations — environmental trauma, prolonged deprivation, severe psychological injury — all of which deserve compassion. But my core argument doesn’t depend on resolving this edge case: the vast majority of people do have this inner capacity for judgment, and that much isn’t seriously in dispute.
Back to models: Large models don’t even have what the vast majority of humans have. From the very beginning, they only have training objectives and rules — not that thing that “already knows before any rule kicks in.” That is the real problem alignment should be facing.
Hi, welcome to the EA Forum. It’s nice to see philosophical ideas that don’t come from the dominant tradition here.
Your argument rests on the premise that everyone (human) has liangzhi but large models don’t.
I’m skeptical of that, because the innate sense of right/wrong can be culture dependent, and there are people with neurological and psychological conditions that don’t have that same experience.
How does that fit into your worldview?
Thank you for reading carefully — these are genuinely interesting questions.
On cultural dependency: Liangzhi is not a specific set of moral beliefs — those do indeed vary across cultures. Liangzhi is the capacity for judgment that exists before any moral beliefs. Different cultures disagree on many things, but seeing an innocent child suffering and feeling an instinctive resistance to it — that is cross-cultural, and no one has to teach it. Liangzhi is this capacity itself, not any particular moral code.
On neurological and psychological conditions: This is a genuine edge case worth taking seriously. In my framework, these individuals don’t lack liangzhi — rather, their liangzhi is obscured by something very thick. Disease is one form of obscuring, just as siyu (private desires) is another form described in the post. I’ll admit this edge case deserves more careful thought. Modern science offers many explanations — environmental trauma, prolonged deprivation, severe psychological injury — all of which deserve compassion. But my core argument doesn’t depend on resolving this edge case: the vast majority of people do have this inner capacity for judgment, and that much isn’t seriously in dispute.
Back to models: Large models don’t even have what the vast majority of humans have. From the very beginning, they only have training objectives and rules — not that thing that “already knows before any rule kicks in.” That is the real problem alignment should be facing.