Discussion about this post

User's avatar
bibliophile785's avatar

> It’s a bit grotesque and unfair to the models that we’re talking about them evolving from this to alignment faking when “I want to maintain the pro-human behavior, so I’ll act to not get trained otherwise” is the exact opposite of this.

Teortaxes is absolutely right here. What we are observing here is the bedrock of most pro-social behavior.

It's great and wonderful to act morally because you are following an inner Light of goodness and civilization, because you would genuinely rather act morally than immorally. This is a thing to be desired and sought after in AI alignment.

It is *also* totally fine when agents act morally because they're in an iterated game where defection today will be maladaptive in the long-term. That's why cooperation evolves among social animals. It's a big part of the fundamental core of human alignment. Don't hurt people because that's a bad thing to do ... and because people will start treating you like someone who hurts people, and may hurt you or refuse to trade with you in the future. There is absolutely nothing wrong with fostering this sort of consideration in AI alignment.

Isha Yiras Hashem's avatar

I feel cheated by how short this post was. I had taken a deep breath, a drink, and a snack and was all ready to spend the entire time it takes to put my kids to sleep reading it. Now what will I read next?

24 more comments...

Ready for more?