Cool, thanks. Of course r and c would depend quite a lot on the task. It’s also an ontology that would diverge significantly from the reality in some important cases. In particular, r is described as a fraction of the automated work, but what are we counting? Is it tokens generated? The human still has to read the judge’s ruling. So we could include tokens processed or whatever? But for a software project, the human has to decide what ze even wants, which can take a lot of thinking, and has to decide some deep architecture stuff, which can also take a lot of thinking; and neither of those are really measurable as a fraction of automated work, if you see what I mean.
Anyway, I think that in some cases the effective r*c constant could be quite high, like .5 or more, leading to less than a 2x speedup. Think for example of generating art. Yes, you could make something ok really fast. But the process of painstakingly going over each bit of the artwork, which could apparently be superfast automated, is actually in many cases an integral part of meditating on what you want, running your fingers (metaphorically or literally) over each square millimeter of the artwork to familiarize yourself with it and with the obstacles and opportunities there.
Cool, thanks. Of course r and c would depend quite a lot on the task. It’s also an ontology that would diverge significantly from the reality in some important cases. In particular, r is described as a fraction of the automated work, but what are we counting? Is it tokens generated? The human still has to read the judge’s ruling. So we could include tokens processed or whatever? But for a software project, the human has to decide what ze even wants, which can take a lot of thinking, and has to decide some deep architecture stuff, which can also take a lot of thinking; and neither of those are really measurable as a fraction of automated work, if you see what I mean.
Anyway, I think that in some cases the effective r*c constant could be quite high, like .5 or more, leading to less than a 2x speedup. Think for example of generating art. Yes, you could make something ok really fast. But the process of painstakingly going over each bit of the artwork, which could apparently be superfast automated, is actually in many cases an integral part of meditating on what you want, running your fingers (metaphorically or literally) over each square millimeter of the artwork to familiarize yourself with it and with the obstacles and opportunities there.
Cf. https://www.lesswrong.com/posts/yCjDGmwQhS7hjEKk5/the-ease-disease