First of all, I think you're claiming a little too much about science from. > the fitness landscape being explored by science is "well behaved" and can be trav

First of all, I think you're claiming a little too much about science from. > the fitness landscape being explored by science is "well behaved" and can be traversed using a simple stepwise or gradient descent function This is false! Non-obvious I know, but it was experimentally shown and holds for a wide class of high dimensional landscapes. It was well explained here (he talks about several novel ideas/insights actually, worth watching): https://www.youtube.com/watch?v=7KCWcx-YIRI https://www.youtube.com/watch?v=7KCWcx-YIRI > There is probably no deterministic "for(;;) { do X }" formula for learning in complex real-world domains This is true in a rather straightforward manner when limited to mathematics, from the halting problem (and some mild computational assumptions about the universe): if no finite algorithm can solve arbitrary instances of the halting problem, clearly no finite algorithm can solve arbitrary problems (that includes algorithms that try to "self-improve", of course). I think this bit of scientific realism is indeed very important to those in the field. > There also seems to be an economic drive to commoditize humans. If learning can be completely systematized, then science can be rendered into an assembly line process... I tend to think of the opposite; the more we're able to replace humans with machines the more power we get to the whatever we want. We just have to make sure this surplus power is made widely available.