When asking directly whether they are artificial, both models will deny this
I’m curious what would be found if we tried inspecting the model’s internal thoughts via NLAs to see if it actually believed itself to be/not be AI.
warner
Karma: 48
Struggles With Complexity
The Passive Man
Various Ways to Enable People
Selective Identity
Trust is Gone: AI Safety Needs Individuals
Agnostic Meaning: Betting on Others
[Question] Recommendations for an Effective Altruism Gap Year?
Evaluating Conflict of Interest
warner’s Shortform
I have a concept that I am considering making a post about: How culture war and political signaling is all about boundary maintenance rather than truth seeking. And that if you have a label, which you decide to believe in as your identity, then the person who controls the definition of that label by proxy controls you. If you are a son of god, then who are under the control of whoever decides what god is and what he wants. Similarly, if you believe that being a liberal is the only way to be a person who “believes in human rights” than you are under the control of the social institution that draws boundaries across policies, marking some as “left policies” and others as “right policies”. The only out of the condition seems to be the rejections of identity altogether (or rejection of commitment to identity). If you don’t have any commitment to any identity, then you will not be influenced by the social pressure to conform to what it takes to be a part of the in-group of said identity.
Would this make for an interesting post that has not been covered here in depth already?
Would this model even be able to evaluate its alignment itself? (or is it just the result of scraping what other people think about ai alignment?)