We could call this "generative adversarial network" (GAN) :)
replyhttps://en.wikipedia.org/wiki/Generative_adversarial_network
This kind of approach would generally still need human guidance, otherwise these models might get stuck in weird niche corners of the problem space that would not be relevant to any real world project.
replyWe could call this "reinforcement learning from human feedback" (RLHF) :)
replyhttps://en.wikipedia.org/wiki/Reinforcement_learning_from_hu...
How do you prevent degenerate strategies? I could trivially give a model a SHA256 hash and ask it to provide the source input.
replyIn class you'd probably want a rule saying at least one LLM should be able to figure out the answer, but in a head-to-head I'm not sure how to solve it.
Maybe make the LLM:s write questions that they can solve (without seeing the question writing context) but not other LLm:s.
replyOn the other hand then maybe a good strategy would be to write questions that the LLM just happen to have in a nich dataset in its training ”what did user5455 say to user6835?”
Nevermind my idea.
deleted
replyWho knows. Maybe Mythos 5 already found a hole in SHA256, so this won't be too hard. :)
replyThat was Fudan I think
reply