15 hours ago by Nick Bostrom

Biological brains are unlikely to be the final stage of intelligence. Machines already have superhuman strength, speed and stamina and one day they will have superhuman intelligence. This is of course not certain to occur it is possible that we will develop some other dangerous technology first that destroys us, or otherwise fall victim to some existential risk.

But assuming that scientific and technological progress continues, human-level machine intelligence is very likely to be developed. And shortly thereafter, superintelligence.

Predicting how long it will take to develop such intelligent machines is difficult. Contrary to what some reviewers of my book seem to believe, I don't have any strong opinion about that matter. (It is as though the only two possible views somebody might hold about the future of artificial intelligence are "machines are stupid and will never live up to the hype!" and "machines are much further advanced than you imagined and true AI is just around the corner!").

A survey of leading researchers in AI suggests that there is a 50% probability that human-level machine intelligence will have been attained by 2050 (defined here as "one that can carry out most human professions at least as well as a typical human"). This doesn't seem entirely crazy. But one should place a lot of uncertainty on both sides of this: it could happen much sooner or very much later.

Exactly how we will get there is also still shrouded in mystery. There are several paths of development that should get there eventually, but we don't know which of them will get there first.

Biological inspiration

We do have an actual example of generally intelligent system the human brain and one obvious idea is to proceed by trying to work out how this system does the trick. A full understanding of the brain is a very long way off, but it might be possible to glean enough of the basic computational principles that the brain uses to enable programmers to adapt them for use in computers without undue worry about getting all the messy biological details right.

We already know a few things about the working of the human brain: it is a neural network, it learns through reinforcement learning, it has a hierarchical structure to deal with perceptions and so forth. Perhaps there are a few more basic principles that we still need to discover and that would then enable somebody to clobber together some form of "neuromorphic AI": one with elements cribbed from biology but implemented in a way that is not fully biologically realistic.

Pure mathematics

Link:

Five ways the superintelligence revolution might happen