zlacker

If this is the case, then just run it X times till error rate drops near 0. AGI solved.

>>nextwo+(OP)
This is called (Algorithmic) Convergence; does the model stably converge upon one answer which it believes is most correct? After how much resources and time?

Convergence (evolutionary computing) https://en.wikipedia.org/wiki/Convergence_(evolutionary_comp...

Convergence (disambiguation) > Science, technology, and mathematics https://en.wikipedia.org/wiki/Convergence#Science,_technolog...

>>nextwo+(OP)
I don’t think it would solve AGI, but having multiple models arguing with each other seems sort of similar to how we work things out when we’re thinking hard, right? Consider a hypothesis, argue for or against it in your head.

>>nextwo+(OP)
As the paper suggested, the LLM cannot identify their own mistakes yet though. And they can only fix their mistakes if the mistake location is given.

>>ming03+Qc
They would fix a "mistake" even if they were given location where there is none.