Superintelligence cannot be contained: Lessons from Computability Theory | Hacker News Reader

Superintelligence cannot be contained: Lessons from Computability Theory

arxiv.org

29 pointsfforflo14 comments

This seems like a very big, pretty unreasonable assumption: "Assuming that a superintelligence will contain a program that includes all the programs that can be executed by a universal Turing machine on input potentially as complex as the state of the world"

This assumption is essentially assuming that "superintelligence" means "god-like intelligence" (god-like because it can compute anything about the world), which is not what the first AGIs that are even significantly smarter than humans would look like.

Indeed, the argument seems a bit nuts. I think a more solid argument for superintelligence not being contained is that human minds will merge with it given the chance so it'll get the same rights as the rest of us.

I think the author has the unstated assumption that "humans will continue to work on smarter and better machines until such a god-like machine is eventually constructed, OR that machines will be created that can self-improve until this state is reached." It doesn't seem unreasonable to me -- the alternative is that at some point in the future, we STOP making machines smarter for good.

There is no reason to believe it is even possible to even create such a God-like intelligence. Even harnessing the energy of the sun would not give you enough energy to compute anything about the universe. We also have every reason to believe that even if this were possible, it would only happen in the far future (e.g. hundreds of years).

Only hundreds?

That's not that many lifetimes.

I was just pulling a number out of thin air. We're in scifi speculation territory so the numbers are all made up.

Fun post. There's some fluff on the way but the central idea that the general function Harm(R, D) is equivalent to the halting problem seems almost trivially obvious once you ask the question. (So much for the laws of robotics.)

My question is whether this approach even models the issue correctly. We humans have problems just defining what terms like happiness, harm, and other general concepts even mean. There's no reason to think that AIs will be any better, whether assisted by us or through their own methods of discovery.

Many successful systems humans have developed for reducing harm take the approach of building up rules for specific cases rather than trying to deduce the cases and their solutions from general principles. Every legal system I can think of works this way and they have by and large worked pretty well. The successful ones have a method for adding to case law in a systematic way, ideally without going to war in the process.

So maybe we should keep the lawyers around after all?

I especially like the story at the end:

> More terrible than either of these tales is the fable of the monkey’s paw[12], written by W. W. Jacobs, an English writer of the beginning of the [20th] century. A retired English working-man is sitting at his table with his wife and a friend, a returned British sergeant-major from India. The sergeant-major shows his hosts an amulet in the form of a dried, wizened monkey’s paw... [which has] the power of granting three wishes to each of three people... The last [wish of the first owner] was for death... His friend... wishes to test its powers. His first [wish] is for 200 pounds. Shortly thereafter there is a knock at the door, and an official of the company by which his son is em- ployed enters the room. The father learns that his son has been killed in the machinery, but that the company... wishes to pay the father the sum of 200 pounds... The grief-stricken father makes his second wish -that his son may return- and when there is another knock at the door... something appears... the ghost of the son. The final wish is that the ghost should go away. In these sto- ries the point is that the agencies of magic are literal-minded... The new agencies of the learn- ing machine are also literal-minded. If we pro- gram a machine... and ask for victory and do not know what we mean by it, we shall find the ghost knocking at our door.

Let's throw in a time factor here, and change the way we define our machine: "Assuming that a superintelligence [SI] will contain a program that can efficiently infer/deduce all the programs that can be executed by a universal Turing machine on input potentially as complex as the state of the world". This seems much more reasonable in terms of how thinking gets done.

We see links to the Halting Problem pop up at times when discussing compilers. Now, compilers try to interpret code in the most perfect way possible. But of course, finding the optimal solution is just the Halting Problem in disguise, and instead they opt for a more pragmatic solution. While none are perfect, some of them are also really, really good, though (http://lemire.me/blog/2016/05/23/the-surprising-cleverness-o...).

Now this is just an analogy, but I think the parallel can be drawn back to the SI scenario. We might not have SIs that never cause harm perfectly, but this does not preclude them from being really, really good at it. Because if you admit the definition above, you're also implying that it will efficiently infer/deduce its harm function.

Cute cartoons, but the entire argument rests on the insane assumption that "a superintelligent machine could simulate the behavior of an arbitrary Turing machine on arbitrary input" (pp.4-5) (!).

...thus being itself Turing-complete? Why is this unreasonable?

Because of the (implicit) assumption that it runs efficiently (as in, at most in polynomial time) even on "arbitrarily complex" inputs... Otherwise the "superintelligence" would not be any different from your run-of-the-mill Turing-complete interpreter!

> Because of the (implicit) assumption that it runs efficiently (as in, at most in polynomial time) even on "arbitrarily complex" inputs...

Uhm... I have no idea where did you get this assumption, because it was not expressed in the sentence we're talking about. Even Turing machine doesn't run in polynomial time on arbitrarily complex input with an arbitrary program.

The super AI doesn't need to surpass the Turing machine, it only needs to efficiently emulate it, i.e. interpret its program. If the TM program was inefficient and ran in exponential time, super AI may run in exponential time as well.

> Otherwise the "superintelligence" would not be any different from your run-of-the-mill Turing-complete interpreter!

Oh. You mean that my mind is not any different from my computer? Because I can run Turing machine (with a little help of external memory, like paper). That's something new.

Thank you! I've been saying as much for years - albeit with fewer proofs and from a more paternal bent.

Reply on news.ycombinator.com