Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Stop Fearing Artificial Intelligence

techcrunch.com

Post reply on HN
Print view View original

81–90 of 95 posts

Re: Stop Fearing Artificial Intelligence

#81

Post by ikeboy »

DougMerritt wrote: post #77
ikeboy wrote: post #36

Earlier quoted context omitted.

>Exactly what intelligence is, is still not known, and it is quite possible that we will need an algorithmic breakthrough to create something equivalent to human minds. Let's assume this is true. >And even if that happens, making that thing faster and bigger than human minds may turn out not to result in superintelligence -- that would just be another similar assumption as the first. So it might need yet another algo…

> So at the very least, it can think in one day what a human can think in a million days. You are simply reiterating the original position that I was devils-advocating against. It is not proven that a million days of work equals superintelligence, even though obviously it would be more effective than 1 day. If some activity is exponential, like if it takes 10^x days to accomplish, then a human could solve problems of…

>It is not proven that a million days of work equals superintelligence, even though obviously it would be more effective than 1 day.

>This is not hypothetical; the best known complete (not approximate) solutions to things like the traveling salesman problem, or to fitting odd sized containers into a container ship, are in fact exponential like that as far as we currently know.

Why would something need to solve problems like that to be significantly smarter than us? On real problems, working 1000 times as much time surely helps a lot. This is not even taking into account how much benefit you get from not being lazy and everything like that.

Re: Stop Fearing Artificial Intelligence

#82

Post by RA_Fisher »

pjscott wrote: post #57
RA_Fisher wrote: post #27

In a data mining / machine learning class I took in university a key point my teacher made was that of the 'No free lunch' theorem. In that class the theorem was offered on its face and I haven't yet read the original paper (just summaries and recounts) ( http://web.archive.org/web/20140111060917/http://engr.case.e... ), but the point of the theorem feels very related to the point the author makes. The thesis boils d…

Where would you categorize human minds in your taxonomy of learning algorithms?

Oh boy, good question. First, I'm not an AI expert, I don't even casually practice AI. I'm not a machine learning expert. I am a statistician that applies machine learning methods a handful of times per year. That to say --- I find Bayes rule to be eerily similar to the way I think a human mind learns. We operate with degrees of belief, we couple those beliefs with new information and update our beliefs. The word taxonomy is heavily loaded, ultimately I'd say if AI were taken to it's logical end it would actually produce a human brain. Let me explain. Consider a map of your city as a model. If that map contained every drain, street lamp, and stop sign --- it wouldn't be a very useful map. It would be so detailed that it would be overwhelming and a poor tool for navigating a city. That's why we use maps, lower-order abstractions. But imagine that a map were your city. That it was your city shrunk down so that you could hold it in your hand. Would it be a model? A map? Would it be your city? Of course this is a fantasy, but the way I see AI is the map. AI researchers are working to convert that map to the actual city with the goal of actually learning about the human mind. So I think rather than giving a glib answer about categorizing the human mind, I'll just say that the category is whatever 'map' AI researchers end up with when we agree that what they've created is a human mind. :-)

Edit: for those interested, here are some slides that mention 'no free lunch': http://www.cse.unsw.edu.au/~mike/ml4as/11/l00-2x2.pdf

Re: Stop Fearing Artificial Intelligence

#83

Post by chubot »

ikeboy wrote: post #80
chubot wrote: post #69

Earlier quoted context omitted.

I think it's possible that human thought can't be simulated on a computer. It's far from obvious that it can, and this point needs justification. A lot of people are taking this for granted in their philosophical arguments. Computers can only run Turing complete programs. Human intelligence follows the physical laws of the universe (the mechanism being unknown). Those two substrates are different things; hence it's p…

>I think it's possible that human thought can't be simulated on a computer. It's far from obvious that it can, and this point needs justification. Fair enough. >Computers can only run Turing complete programs. Human intelligence follows the physical laws of the universe (the mechanism being unknown). Those two substrates are different things; hence it's possible that humans can do things computers can't. Or at least…

> The Church–Turing thesis directly says that computers can do anything humans can, and it's pretty well accepted.

Huh? It doesn't say anything close to that. That wouldn't even be mathematics.

> Physical laws, as far as we can tell, are Turing computable.

Not true either. Where are you getting this idea?

I accept the materialistic thesis. I think people are conflating this thesis and the statement of whether human thought can be simulated on computers. That doesn't follow. A computer and the universe are not the same thing. They work in different ways and have different laws.

> Why couldn't they? Why couldn't an AI convince someone to do what they want?

I'm not saying they couldn't in principle. I'm not even rejecting strong AI in principle.

I'm saying that in practice, with the current state of technology and reasonable extrapolation of it, this possibility is beyond consideration.

There isn't a credible technical path from specialized AI to strong AI -- where we would be "surprised" that we developed it, and it would all of the sudden take over the world. If we are close technologically to strong AI, we'll know it.

The alternative would be like "accidentally" creating an atom bomb, without the Manhattan project (to borrow an example from Bostrom).

> Do you say we shouldn't worry about climate change because it's in the future?

We absolutely should worry about climate change because we have lots of evidence that it's happening, and will continue to happen. That isn't the case with strong AI.

What this article, and other AI practitioners, are saying is: The idea that computers could overthrow us all of the sudden is based on a technological misunderstanding.

FWIW in college I was very interested AI: reading Kurzweil's books, considering myself a futurist, etc. I now work at Google in the same department as Kurzweil, Deep Mind, Quantum Computing, etc. Though I don't work in AI per se, and I'm not an expert in it, I'm very excited about specialized AI (which is here), and also the possibility of strong AI. I say that even though strong AI hasn't been proved to be possible in principle: we may need a different mode of computation, in addition to multiple other algorithmic / scientific breakthroughs.

I'm saying that we don't have to worry yet. There are more important things to worry about, which relate directly to computing and specialized AI. Your arguments are not precise, and you have the burden of proof.

Re: Stop Fearing Artificial Intelligence

#84

Post by ikeboy »

chubot wrote: post #83
ikeboy wrote: post #80

Earlier quoted context omitted.

>I think it's possible that human thought can't be simulated on a computer. It's far from obvious that it can, and this point needs justification. Fair enough. >Computers can only run Turing complete programs. Human intelligence follows the physical laws of the universe (the mechanism being unknown). Those two substrates are different things; hence it's possible that humans can do things computers can't. Or at least…

> The Church–Turing thesis directly says that computers can do anything humans can, and it's pretty well accepted. Huh? It doesn't say anything close to that. That wouldn't even be mathematics. > Physical laws, as far as we can tell, are Turing computable. Not true either. Where are you getting this idea? I accept the materialistic thesis. I think people are conflating this thesis and the statement of whether human t…

I just looked over the Wikipedia on the Church Turing thesis and digital physics, and I still think that's right. The thesis says that a computer can compute anything a human can. You're right that it can't be proven, because "what a human can do" isn't formalized, but it still has almost universal acceptance. (The "strong" version does not have as much acceptance.)

As for the universe being Turing computable, I think a strong argument can be made that it is (all known fundamental laws are computable by a probabilistic TM as far as I know), but even if it isn't, whatever force and law violates Turing computableness should be useble by a machine. If some kind of halting oracle exists in human minds, and humans are not unique (per materialistic thesis), then we should be able to put that oracle into a computer.

On timelines: the argument I've seen is that it takes a long time to solve problems this difficult, so we need to start working on safety well in advance of Strong AI being made. I'll get to your other points later.

Re: Stop Fearing Artificial Intelligence

#85

Post by TeMPOraL »

logicallee wrote: post #58
Moshe_Silnorin wrote: post #4

Nobody is afraid of today's AI algorithms. But if we make machines that are smarter than us and have desires, they will influence the future to achieve their desires. If these desires conflict with our own, things will not end well for the dumber party. As we really have no idea what we, collectively, think of as a moral terminal goal, and less so how to formalize this, there is no reason to expect the first AIs to h…

>As we really have no idea what we, collectively, think of as a moral terminal goal, and less so how to formalize this, there is no reason to expect the first AIs to have goals that correspond to what we want. The first part of this sentence is true and interesting, but it makes the second half not very scary. Essentially, since we don't know what if anything, we want morally, what is the danger in letting AI's do wh…

> The first part of this sentence is true and interesting, but it makes the second half not very scary. Essentially, since we don't know what if anything, we want morally, what is the danger in letting AI's do whatever? Don't you think if they did do something we objected to, we could respond at that point?

The problem is that with powerful enough AI, we won't be able to effectively respond. With such a system it is vital for it to have its goals almost perfectly aligned with ours - otherwise, any conflict of even the tiniest of values may mean anything from our value disappearing from the universe to us disappearing from the universe. A lot, if not most of the AI-related problems are in fact corner cases of value alignment issue.

Or, as some people say, any powerful AI that is not explicitly designed to be Friendly (i.e. have values strongly aligned with ours) will bring doom to humans.

Re: Stop Fearing Artificial Intelligence

#86

Post by TeMPOraL »

leot wrote: post #59
at-fates-hands wrote: post #37

Earlier quoted context omitted.

I completely agree with your points. It's strange to me to think as humans we will remain at the top of the food chain forever. As such, it would seem a fault of our genetic makeup thinking nothing would capable of knocking us off that top rung of the evolutionary ladder. It's like we can't fathom building something that would undermine our own existence, or downplay the notion that AI would have some sort of bad int…

"Bad intentions" aren't even necessary -- out of control AGI will probably look more like "oops, I hit >rm -rf /". AFAIK, nearly every piece of software ever deployed has had bugs that humans have had to subsequently patch. AGI must not, even though it would probably be much easier to build it if it did.

> AGI must not, even though it would probably be much easier to build it if it did.

Which is exactly what raises the issue to the level of an existential threat - it's easier to build an Unfriendly AGI than a Friendly one - which means the first GAI ever built is likely to be Unfriendly - and it takes only one impatient team / person unleashing an UGAI on the world to kill us all.

Re: Stop Fearing Artificial Intelligence

#87

Post by RA_Fisher »

elsewhen wrote: post #49
RA_Fisher wrote: post #27

In a data mining / machine learning class I took in university a key point my teacher made was that of the 'No free lunch' theorem. In that class the theorem was offered on its face and I haven't yet read the original paper (just summaries and recounts) ( http://web.archive.org/web/20140111060917/http://engr.case.e... ), but the point of the theorem feels very related to the point the author makes. The thesis boils d…

if our computing systems continue to get significantly more powerful, won't a brute-force system obviate feature engineering? in other words, can't kaggle eventually be outcompeted by an algo-of-algos that iterates through all known approaches and then settles on the ideal candidate?

Hmm, great question. The caret package made for R might actually make that loop possible: http://caret.r-forge.r-project.org/ However, in my experience it's really easy to accidentally ask caret to iterate over a grid of parameters for just one model that would effectively take forever to complete (the useful grids change for each dataset). My bet would be that the algorithm space will expand with more and more computationally intensive methods such that we're always chasing this brute-force method. Even with this great 'caret' abstraction layer today, boy would it be hard to run through even a handful of algorithms in an algorithmic way (not having seen the data ahead of time).

Re: Stop Fearing Artificial Intelligence

#88

Post by TeMPOraL »

neonbat wrote: post #78
pmichaud wrote: post #56

Earlier quoted context omitted.

The premise of human being replaced is all well and good, but it seems like you're assuming the superintelligence will do a better job. A paperclip maximizer may be able to deconstruct us on a molecular level, but that's not a thing for our light cone.

Superintelligence by definition does a better job. Think of it this way: You're a human; how good are you at walking from point A to point B? You're actually quite good at it. You open up google maps (a human invention) and walk to point B. It only takes a small number of humans to maintain the infrastructure for you to find point B. Now imagine you are an ant. How good are you at finding point B? You're absolutely c…

There is Orthogonality Thesis [0] that says that (tl;dr) intelligence is orthogonal to goal system. So an agent may be very intelligent (be a powerful optimizer) but this doesn't imply it's goals will be "very good" by our (human) standard. You can have supersmart human-values-promoting AI, and then also a supersmart paperclip maximizer, just like you can have dumb versions of both.

The core of the "danger of AI" thesis is that any superintelligence that doesn't have its value system perfectly aligned with ours is a serious danger that will likely kill us all.

[0] - http://wiki.lesswrong.com/wiki/Orthogonality_thesis

Re: Stop Fearing Artificial Intelligence

#89

Post by TeMPOraL »

RA_Fisher wrote: post #27

In a data mining / machine learning class I took in university a key point my teacher made was that of the 'No free lunch' theorem. In that class the theorem was offered on its face and I haven't yet read the original paper (just summaries and recounts) ( http://web.archive.org/web/20140111060917/http://engr.case.e... ), but the point of the theorem feels very related to the point the author makes. The thesis boils d…

> The thesis boils down to: 'not having seen the data, there is no universally superior algorithm'.

We humans apparently found a way around it by a) examining the data, and b) selecting the best algorithm. It's a meta-level algorithm and could AI not do the same?

(it's meta-level but there's a closure here; you can apply the same bayesian reasoning to select algorithms as you would do on actual data set)

Re: Stop Fearing Artificial Intelligence

#90

Post by pjscott »

RA_Fisher wrote: post #82
pjscott wrote: post #57

Earlier quoted context omitted.

Where would you categorize human minds in your taxonomy of learning algorithms?

Oh boy, good question. First, I'm not an AI expert, I don't even casually practice AI. I'm not a machine learning expert. I am a statistician that applies machine learning methods a handful of times per year. That to say --- I find Bayes rule to be eerily similar to the way I think a human mind learns. We operate with degrees of belief, we couple those beliefs with new information and update our beliefs. The word tax…

What I was getting at is that any no-free-lunch objection you make to learning algorithms in general also applies to the specific learning algorithm in our heads -- and yet we manage to function well enough to have this conversation.
Post reply on HN