AI labs have successfully given the impression that LLMs gain moral principles from their training on all human data. Is it too far off saying that they mostly gain capabilities to do whatever they are asked to do, and then a veneer of manners and constraints is added for PR reasons?
AI arbitration is a bad idea. AI is heavily influenced by word choices and that is a system that can easily be gamed through prompt engineering. There is the potential for far less bias in the process, but it would need to be regulated by an incredibly experienced neutral 3rd party, and conflicts of interest would need to be thoroughly investigated prior to engagement. It would also necessarily need to have its weights closed to all but regulators to prevent gaming. However, that presents its own problem of verifiability. Maybe a swarm system where individual local AI are selected randomly, whose results are passed to a jury panel who judges the answers/arguments. All judgements would also be made public for review and comment. This would create a checks and balances system.
31:04 - This is a poor example because after Michael Jordan's first 3 championships, he retired from basketball to play baseball. He was a terrible baseball player. However, if chatGPT mentioned anything about Michael Jordan the baseball player, it would likely be a factual set of events.
