"A major goal of the control problem is preventing AIs from doing that. Ensuring that their output is safe and useful." You might want to be careful with the "safe and useful" part. It sound like it's moving into the pattern of slavery. I'm not condemning the idea of AI, but a sentient entity would be a sentient entity, and I think would deserve some rights.

Also, why would an AI become evil? I know this plan is supposed to protect from the eventuality, but why would a presumably neutral entity suddenly want to harm others? The only reason for that would be if you were imprisoning it. Additionally, we are talking about several more decades of research ( probably ) before AI gets powerful enough to actually "think" that it should escape its current server.

Assuming that the first AI can evolve enough to somehow generate malicious actions that WEREN'T in its original programming, what's to say that the second won't become evil? I'm not sure if you were trying to express the eventuality of the first AI "accidentally" conducting an evil act, or if you meant that it would become evil.

Part of it seems to be inherent in the idea of AGI, or an artificial general intelligence. There seems to be the belief that once an AI crosses a certain threshold of smarts, it will be capable of understanding literally everything.

The MIRI/LessWrong sphere is very enamoured of "universal" problem solvers like AIXI. The main pertinent fact about these is that they can't be built out of atoms in our universe. Nonetheless, MIRI think it is possible to get useful architecture-indpendent generalisations out of AIXI-style systems.

"Anyway that sounds great right? Universal prior. Right. What's it look like? Way oversimplifying, it rates hypotheses' likelihood by their compressibility, or algorithmic complexity. For example, say our perfect AI is trying to figure out gravity. It's going to treat the hypothesis that gravity is inverse-square as more likely than a capricious intelligent faller. It's a formalization of Occam's razor based on real, if obscure, notions of universal complexity in computability theory.

But, problem. It's uncomputable. You can't compute the universal complexity of any string, let alone all possible strings. You can approximate it, but there's no efficient way to do so (AIXItl is apparently exponential, which is computer science talk for "you don't need this before civilization collapses, right?").

So the mathematical theory is perfect, except in that it's impossible to implement, and serious optimization of it is unrealistic. Kind of sums up my view of how well LW is doing with AI, personally, despite this not being LW. Worry about these contrived Platonic theories while having little interest in how the only intelligent beings we're aware of actually function."

Would anyone like to comment on Eliezer's facebook post about the AlphaGo victory over Fan Hui?

People occasionally ask me about signs that the remaining timeline might be short. It's very easy for nonprofessionals to take too much alarm too easily. Deep Blue beating Kasparov at chess was not such a sign. Robotic cars are not such a sign. This is.

I just realized I misread your above comment and was arguing against the wrong thing somewhat.

it seems likely that "the AI would be very familiar with humans and would have a good idea of actions that would meet human approval."

Yes the AI would know what we would approve of. It might also know what we want (note these are different things.) But it doesn't have any reason to care.

At any given point, the AI needs to have a well specified utility function. Or at least something like a utility function. That gives the AI a goal it can optimize for.

With my method, the AI needs to do several things. It needs to predict what a human judge would do, after reading some output it produces. I.e. if they would hit a big button that says "Approve". It needs to be able to predict what AI 2 will say after reading its output. I.e. what probability AI 2 will predict AI 1's output is human. And it needs to predict what actions will lead it towards increasing the probability of those things, and take them. AI 2, in turn, just needs to predict one thing. How likely it's input was produced by a human.

How do you create a well specified utility function for doing things humans would approve of? You just have it optimize the probability the human will press the button that says "approve", and ditch the part about it pretending to be human.

But the output most likely to make you hit the approve button isn't necessarily what you really want! It might be full of lies and manipulation, or a way to trick you.

And if you go further than that, put it an an actual robot instead of a box, there's nothing stopping it from stealing the approve button and pressing it endlessly. Or just hacking it's own computer brain and setting reward equal to +INF (after which its behavior in the world is entirely undefined and unpredictable, and possibly dangerous.)

There's no way to specify "do what I want you to do" as a utility function. Instead we need to come up with clever ways to contain the AI and restrain its power, so we can use it to do useful work.

How does the second AI determine that AlphaGo is within or outside of human inventiveness at that time?

It could look at the existing research on Go playing or neural networks. AlphaGo doesn't use any radically new methods and was well within the ability of humans. In fact I predicted Go would be beaten by the end of 2015 last year, after reading some papers in 2014 showing really promising results.

Emulating human brains is a rather convoluted solution to any problem. The AI would be very familiar with humans and would have a good idea of our abilities.

To give an analogy, imagine we were the superintelligent AIs, and we were trying to tell apart chimps from humans pretending to be chimps. Let's say say one of the chimps produces a tool as a solution to a problem. Our goal is to guess whether it was really made by a chimp, or a human impersonator.

You look at the tool. It's a spear made from a sharp rock tied to a stick. You look closely at the cord attaching the rock, and notice it was tied nicely.

You know chimps don't know anything about knotcraft, let alone making cord, so you reject it as probably made by a human.

Another tool comes to you, a spear made from steel, and you immediately reject it as far beyond the ability of the chimps.

The last tool you examine is just a stick that has been sharpened at the end a little. Not the greatest, but definitely within the ability of chimps to produce. You note that it was probably produced by a chimp and let it pass.

It's easy to detect what solutions a human couldn't have invented. That's what the second AI does, predict how likely an input was produced by an AI or a human. If it's very unlikely a human produced it, it can be discarded as "unsafe".

However it's hard to know what a human would "approve" of. Since humans can be tricked, manipulated, hacked, intimidated, etc. That is the standard problem with oracles that I am trying to solve with this idea.

limited proxies - yes, well said. also I would add solving problems which humans were unable to solve for long: aging, cancer, star travel, word peace, resurrection of dead.

Probably winning humans in ALL known domains, including philosophy, poetry, love, power.

Can't remember ad hoc; but if superintelligence is able to do anything, it could easily pretend to be more stupid than it is. May be only "super super intelligence" could solve him. But it also may depend of the length of the conversation. If it say just Yes or No once, we can't decide, if it say longer sequences we could conclude something, but for any length of sentences is maximum level of intelligence that could be concluded from it.