Current AI safety approaches have a structural problem: they ask the same system to both generate outputs and judge whether those outputs are safe. RLHF trains a model to produce outputs that a reward model likes. Constitutional AI has the model critique its own outputs against principles. In both cases,...
My eyes are burning from the strain of how aggressively I am not-crying right now. For as long as I can remember I have been the odd one out. I never cared about what everyone cared about, or at least not in the "normal way" that people were expected to...
Current AI safety approaches have a structural problem: they ask the same system to both generate outputs and judge whether those outputs are safe. RLHF trains a model to produce outputs that a reward model likes. Constitutional AI has the model critique its own outputs against principles. In both cases, the system interpreting the constraints is the system being constrained. This is like asking a defendant to also serve as judge and jury.
If the objective is human level intelligence, then it would be prudent to define how human intelligence operates. The human mind makes deliberations on "rightness" or "wrongness" at multiple stages in the thought process, and at different levels of consciousness,... (read 1198 more words →)
My eyes are burning from the strain of how aggressively I am not-crying right now.
For as long as I can remember I have been the odd one out. I never cared about what everyone cared about, or at least not in the "normal way" that people were expected to care about things. Sports, pop culture, music, whatever that week's trending vice happened to be, even literature was really only of one specific use to me: information.
Because I care about being Correct. Not being Right, which is a different dragon entirely, but to know What and Why and How everything is or was. I have this indomitable curiosity that insists, if I am... (read 200 more words →)