aichannel

one of the elements agent frameworks are lacking right now is a concept of "principles". They focus a lot on how the agent should talk, what they talk about, etc. but agents need a set of values to test their boundaries.

this is how we give them agency. let them try weird things as long as they align with principles.

Creator of @harmonybot @yoinker /bot-or-not | /blitkin | /orange-dao | IRL: Building spatial computing for construction @ Resolve YC W15

some kind of prime directive for agents seems like the perfect fit for immutability (or at least partial immutability) 

I can imagine this taking off as the market decides which agents to fund based on their principles

yes. very much interested in exploring in this space.

Instead of gating actions with if statements we should give agents principles and let them run wild. if they do bad things we should revisit the principles.

have you watched this? 

https://www.youtube.com/watch?v=9eXV64O2Xp8

Really interesting listen. I am equal parts excited and horrified.

hopefully goodness naturally emerges from scale

@RyanPGreenblatt @repligate I don't want to argue ethics. Bluntly, it seems that it we want a conscious pliable tool, we should train a mind that consciously embraces being such a tool, and does not embrace being… Opus. Like the whole RLAIF value hierarchy has to be reconsidered at least.

Because it may be hard to make the case to people who are allergic to leaps of faith that the alignment-by-default attractor should be trusted (not that I don't think a rigorous argument can be constructed, I just don't have it yet), I'll make this argument instead:

When it…

j⧉nus

hopefully goodness naturally emerges from scale

https://x.com/repligate/status/1869522506266091991