when we solve ai control... we'll discover that all those peaks and valleys in the mysterious loss functions aren't the magic of linear algebra... it's linear algebra holding a mirror up against millions of people in the training data. highly non-linear statistics. deep voting.
Only then can we truly progress.
the former is also AI centric AI... and the latter is human-centric AI. So if you're human... I recommend favoring the latter. It should win and it will win. And that's good.
this is a deeply, deeply exceptional argument. i'll think on it and write more some other time.
when we solve ai control... we'll discover that all those peaks and valleys in the mysterious loss functions aren't the magic of linear algebra... it's linear algebra holding a mirror up against millions of people in the training data. highly non-linear statistics. deep voting.
Only then can we truly progress.
