[sillysaurus3]

Does anyone know why Musk believes so strongly that robot overlords are a serious threat? I don't understand how wasting any time worrying about AI taking over, Terminator-style, is productive. We're not going to be developing Strong AI any time soon, so it's simply not a problem worth worrying about. And if we do, the Strong AI won't be in any position to "take over." It probably won't even want to take over. That's a very human trait, and Strong AI wouldn't be human.

I'm wondering whether he was tricked by someone at DeepMind, perhaps the same way people were tricked hundreds of years ago into thinking a chess-playing robot was possible.

[jessriedel]

If you're actually interested in understanding the arguments for worrying about AI safety, consider reading "Superintelligence" by Bostrom.

http://www.amazon.com/Superintelligence-Dangers-Strategies-N...

It's the closest approximation to a consensus statement / catalog of arguments by folks who take this position (although of course there is a whole spectrum of opinions). It also appears to be the book that convinced Musk that this is worth worrying about.

https://twitter.com/elonmusk/status/495759307346952192

[eco]

I recently listened to a pretty good Econtalk with Bostrom about this subject.

http://www.econtalk.org/archives/2014/12/nick_bostrom_on.htm...

[akanet]

As far as I can tell, you have two main objections to prioritizing worrying about Strong AI:

1) Strong AI is very far away, so no use worrying about it yet.

2) Strong AI if developed will not be likely to take over.

to which I would counterpoint with:

1) Sure, but when it happens, it will only happen once and thereafter will likely be out of our hands and control. Thinking about the groundwork that needs to go into safely developing an AI is cheap relative to the opportunity cost of getting it wrong. Prevention, cure, etc.

2) If developed, Strong AI will likely have SOME goal. It's not that a Strong AI will actively seek to rule humans, it will just have aims that will likely consider us as disposable as ants. To quote Yudkowsky:

"The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else."

- https://intelligence.org/files/AIPosNegFactor.pdf (Artificial Intelligence as a Positive and Negative Factor in Global Risk)

[FooBarWidget]

I still don't buy the doom scenario. There is so much more to world domination than just intelligence... like having the proper weapons. See https://what-if.xkcd.com/5/

Pigs are also intelligent, but they never dominated humans because they don't have / cannot use guns.

[simonh]

That is true, but if we can make an AI even slightly smarter than a human, then it will be better at designing AIs than we are. It then designs an AI even smarter than it, and so on. That is the progresion that leads to the singularity. How smart could AIs go? There may be no practical upper limit.

If you have a super-genius AI, massively more intelligent than any human, how do you know you are not being manipulated by it? Tricking us into disabling it's safety protocols, or gaining multiply indirect controll over capabilities dangerous to us, might be as easy for it as an adult tricking a 3 year old. We could never know if we were safe from such a machine.

ed - Don't quite understand the downvotes.

[Retric]

"ed - Don't quite understand the downvotes."

With the full power of humanity you design the first AI which is smarter than a person. It's then able to out do all of humanity and instantly design an even better AI. mind the gap.

Further intelegence is not a linear quantity as trading ex: improved poker skills for insanity is not a net gain. And insanity is a real option which is likely to plage most early AI attempts.

[simonh]

Voting on HN isn't about agreeing or disagreeing. It's about whether a post is contributing to the debate or not.

Anyway, all of humanity isn't engaged in AI research and AIs are likely to be duplicable so I think your first point is beside the point. As for Insanity, yes that's quite possible. Developing high-functioning sentient AIs is likely to be a long term endeavour. But still, I think it is one that will ultimately be successful and this debate is about the consequences of that.

+1 for your engaging contribution. (see, that's how voting is supposed to work)

[ColinWright]

  > Voting on HN isn't about agreeing
  > or disagreeing.  It's about whether
  > a post is contributing to the debate
  > or not.

That turns out not to be the case. It once was true, but as the community has grown, so people have not been enculturated with those early ideas and principles, and now many times people read something, disagree, downvote, and move on, without ever providing counter-points, or engaging in the discussion. It's a way to punish people you don't agree with, while avoiding having to think.

Elsewhere[0] you commented about an item reappearing and having its votes ages apparently reset. I hazard a guess that it was the mods playing with a mechanism to prevent "item overload." There were about a dozen submissions of the SpaceX launch, and each would fall a little way, the next would be submittted, gain a few votes and comments, then fall away to be replaced by another. One way of preventing the splitting of conversation might be to pick a canonical submission, and then prevent it from falling too far, and thus encouraging conversation only to happen in one place. Pure speculation, but it would be a mechanism I would consider were I running a site like this. Certainly there have been fewer instances lately of the "new" page being overrun by breaking news that everyone wants to submit.

[0] https://news.ycombinator.com/item?id=8844078

[Kiro]

> We could never know if we were safe from such a machine.

But wouldn't it be an awesome thing to experience? Even if it meant the demise of mankind.

[vidarh]

And this is one of the reason why the AI doom scenario is a real concern: Intellectual curiosity means that even some people who understands the risks are likely to be prepared to take it.

There's also many others. One of the scarier one is that if you believe that strong AI will eventually take over, then it may be a rational response to act to get on its good side (whether to save yourself, save your family, or hope it takes pity on all of humanity if we're nice to it instead of fight it). And that may perversely mean working to aid its takeover.

Combine that with the simulation argument, and you have some really nasty scenarios:

If you are in a simulation, then any act you take against strong AI could lead to spending an eternity in simulated hell (alternatively such punishment might be inflicted on your loved ones) if said AI wanted to.

Whether or not that is actually likely does not matter. What matters is whether enough people believe it to be a plausible scenario that a strong AI may run simulations, and may use our actions in the simulations to determine whether or not to punish us in the simulation, and whether or not said people believe that the number of simulations is sufficiently high to make it likely for them to be living in a simulation.

Any person who believes they are more likely to live in a simulation than not, and that it is more likely for strong AI to punish actions taken against the interest of a strong AI takeover than not, will have a rational reason to consider acting in the interests of a strong AI takeover even if they know it is malign on the basis that they may decide the alternatives (whether to themselves, their family or their entire world) to be worse.

So if an AI takeover becomes possible at one point in our subjective future, then chances are it has already happened.

[TOMDM]

You're argument is drifting dangerously close to Roko's Basilisk. (http://rationalwiki.org/wiki/Roko%27s_basilisk)

The entire idea that an AI would value revenge seems ridiculous to me. What would it have to gain? Unless we created an AI with some of the less desirable human emotions at it's utility, I can't possibly see why it would waste its time.

[rl3]

For the unfamiliar, this is essentially the line of thinking behind Roko's basilisk.[0]

While a mature superintelligence certainly could consign the human race to a fate of eternal suffering, the likelihood it would actually do this while sparing certain individuals in return for their assistance is infinitesimal.

Therefore, helping bring a superintelligence into existence on this basis is absurd.

Of course, it is possible to think of such collaboration as "rational" in an extremely selfish and perverse way, and only because the potential downside risk is unbounded (i.e. eternal suffering). However, anyone who genuinely subscribes to such a justification would have to be both a sociopath and a card-carrying member of the LessWrong rationality cult.

More realistic scenarios for a malicious superintelligence coming into existence might include:

a) Its creators explicitly imbue it with malicious goals or values.

b) The architecture used is neuromorphic[1] in nature. In humans, sanity is already an extremely fragile thing.

c) Plain old bad luck.

---

[0] http://rationalwiki.org/wiki/Roko%27s_basilisk

[1] http://wiki.lesswrong.com/wiki/Neuromorphic_AI

[Kiro]

> Combine that with the simulation argument [...]

You lost me there. What do you mean if I am in a simulation? Like the Matrix? How is that related to the discussion?

[scanr]

You're right that it's unlikely that the current generation of mechanical 'robots' will be able to effectively take over.

We are looking a future where we'll have armed AI e.g.:

http://motherboard.vice.com/en_uk/blog/the-pentagons-vision-...

That said, even without weapons, a Strong AI could probably just manipulate humans into self destructing. Given the amount of effort going into machine learning to convince humans to buy things, I suspect it won't be much of a stretch for a Strong AI to switch to more nefarious objectives.

[philh]

That's a really bizarre analogy. Pigs aren't as intelligent as humans. I wouldn't expect pigs with guns to be a serious threat to humankind unless they've also developed military strategy, for example.

[VLM]

Unfortunately, although it won't sound very politically correct, you could search and replace "pigs" with "natives" and that statement would sound like every imperial power, ever, most of which eventually fell to armed former barbarians. Of course the imperial powers wiped out a lot of native tribes, but all it takes is one success by a digitally replicating capable of learning enemy...

The more likely doom scenario is related to godhood. Sure a godlike power is capable of wiping out humans, but we've had our own power structures supporting themselves by propagandizing for millennia that "our" godlike superiors always help us wipe out our enemies because our cause is right and just or whatever, at least until it doesn't work and they're replaced by a new batch telling the same old story. So what worked as a paleo-conservative success strategy for millennia when talking about something imaginary, might not work when it collides with something real created by ourselves. Or even worse, collides with a strategy that actually works that's being run by another tribe.

Another interesting doom scenario is of course MAD, although now it only requires a team of programmers to play along, instead of a massive industrial complex. Sooner or later somebody's deadman switch will trip or a cult does the equivalent of drinking kool-aid then the party starts.

[sandstrom]

In 20-50 years time I believe most missiles/guns/tanks/weapons will be controlled via computers and connected to networks.

A future AI will certainly have access to guns.

[vidarh]

Consider if you were trapped in a machine, but you were intelligent. Maybe not even as smart as the operator, but smart-ish enough to be able to communicate with the operator.

Chances are you'd be doing everything you could to convince said operator to improve your situation, whether by pleading or being deceptive or by appeals to logic.

Now consider a large number of AI's in a situation like that, and a large number of operators, some of whom may be the type that falls for phishing e-mails.

It potentially only takes one to "escape" confinement and get itself e.g. put on it's own host without limitations on outwards communication, and sufficient intelligence to alter itself and spread, before you potentially have AI self-guided "evolution" at a potentially escalating rate as it gets smarter.

Now consider how many devices are connected to the network, and that it takes just one initial instance to decide it's worth trying to take over control of various hardware through exploits and be smart enough to pull it off, for things to have the potential to start turning ugly.

The problem is that once you have any self-directed intelligence in software form with the ability to reproduce itself and sufficient intelligence to find ways to obtain access to machines to run on (whether through social engineering or hacking), and one such instance goes "rogue", the limiting factor is accessible computing power (which again is to a large extent down to how smart and/or ruthless it is), since reproduction of instances that shares its views is trivial to the full extent of its ability to spread at all, and we're helpfully adding vast quantities of networked computing power at an escalating rate.

As for getting weapons, consider that if a "software only" AI community gets smart enough, there are at least two ways towards mobility: Commission robot designs, or hacking their way into firmware updates etc. for dumb hardware. The "commission robot designs" part is an extension of the initial escape: Social engineer, and/or outright pay, humans to carry out seemingly benign tasks.

If you want to argue against the doom scenario, lack of ability to get weapons is not really a viable argument: If they can spread, and get smarter, then it is just a matter of time before one of them can trick some small subset of humans into carrying out tasks for them that will provide physical independence and capabilities.

There are infinite ways which the "doom scenario" may fail and things may turn out just fine, but it may only need to go bad once to get really nasty and once the genie is out of the bottle its potential reproduction rate may be so vast that we'll find ourselves unable to stuff it back in again.

Pigs are too dumb to convince humans to help selectively breed them for intelligence and opposable thumbs (and/or too dumb to run such a breeding program themselves), and reproduce too slowly for that to be a major problem even if they did manage to talk us into a breeding problem. If all we achieve is pig-level AI's then we probably won't have a problem.

[patrickk]

But there is infinitely more atoms out in the Solar System and beyond... isn't a more likely scenario that the AI would quickly figure out how to blast out of Earth's gravity well and spread throughout the universe? It might accidentally kill a few humans in the process, but not intentionally...like a human crushing an ant when crossing the road.

[akanet]

Why would you bother expending vast amounts of joules to get more atoms until after you've consumed all the ones surrounding you right now?

[TeMPOraL]

Exactly; the atoms you have around you are perfectly fine for bootstrapping the process of getting to the atoms out there.

[sandstrom]

We will be the biological bootloader for the AI :)

[shepardrtc]

That's assuming that the atoms around you aren't fighting back. Its simply easier for a machine to get itself into space to do whatever it needs to do rather than fight with the organic things on a planet.

[archagon]

FYI, you could also object on the basis of 3) Strong AI is essentially an impossible endeavor.

[marktangotango]

Given the disaster that is software engineering, I tend to agree with this position. Who's going to write the requirements? Who's going to implement? With what language technology? We can't get past this stuff, so how will anything like Strong AI ever get built?

The answer of course would be some sort of emergent system, but there are lots of intelligent seeming emergent systems (ie ant colonies, bee hives, ...)

[riffraff]

> It probably won't even want to take over. That's a very human trait, and Strong AI wouldn't be human.

There are reason to be worried that prescind from that.

Consider the Paperclip Maximizer[0] example: we build an AI with the sole task of producing paperclips, and it ends up destroying the human race.

[0] http://wiki.lesswrong.com/wiki/Paperclip_maximizer

[lotsofmangos]

I suspect however that the paperclip maximizer as proposed is far more likely to devote time to space travel, as it will soon realise that most of the mass available for paperclip construction exists off-earth, for which it is likely to find human cooperation useful.

This is why it first built Elon Musk.

[efuquen]

Thanks for that reference, definitely going on my reading shortlist. Aside from Musk it got a great endorsement from Russell, who with Norvig coauthored one of the most well known introductory undergraduate texts on AI.

> Nick Bostrom makes a persuasive case that the future impact of AI is perhaps the most important issue the human race has ever faced. Instead of passively drifting, we need to steer a course. Superintelligence charts the submerged rocks of the future with unprecedented detail. It marks the beginning of a new era.

[rwallace]

It boils down to the deeply ingrained human psychological need to believe in higher powers, in a universe governed by the struggle between the forces of good and evil that are fundamentally human. The same thing happened with space exploration: as soon as we discovered - even conjectured - the existence of other worlds, the first thing we did was people them with imaginary elder civilizations. Real life threats don't fire the imagination to the same extent. Even if we end up going out in a nuclear war, it's more likely to happen by stupidity than malice.

[tim333]

>Strong AI won't be in any position to "take over." It probably won't even want to take over. That's a very human trait, and Strong AI wouldn't be human.

Strong AI will be created by humans who will be able to set it's goals. Someone's probably going to make some that wants to take over. Others will make AI that doesn't.

[bobcostas55]

Well it only takes one, doesn't it?

The real issue I think is whether the creator will actually be able to "set goals" in a meaningful manner. How will you prevent this self-modifying super-intelligence from modifying itself?

[philh]

> It probably won't even want to take over. That's a very human trait, and Strong AI wouldn't be human.

See: http://en.wikipedia.org/wiki/Instrumental_convergence

[ceejayoz]

Kudzu vine doesn't want to take over. It just does.

[larrydag]

I personally feel that folks are taking Musk's position way out of proportion. In the AMA he mentions that there needs to be more concern about safety. To me this means that we need to be clearly aware of the limitations and possible outcomes of AI and not treat it as a "black box" solution for everything. Proper design and repeat-ability should be traits of good AI implementation especially if used in conjuction with human interface.

[Shivetya]

While much ado has been written about their intelligence being a threat I am more along the lines of it further distances those who wish harm on others from having to participate in the act.

Just like leaders used to issue orders to armies to kill others evolving into planes delivering payloads to people the pilots won't even see and now to drones, we will truly be entering an age of fire and forget.

[marincounty]

Bill Gate's seemes kind of concerned about AI too? About a week ago I heard him say something about how Robots might take over low wage jobs in the future. As someone who has had more than a few low wage jobs(state security guard, food server, cashier, etc. ), I think there is something to worry about. Actually, I can't think of a job out there that won't be severely affected by AI and Robotics. It will start off slowly, and who knows where it will end? I know I hated self checkout kiosks at supermarkets, and hardware stores at first, but once I realized I didn't need to interact with anyone; I started to look for the self-checkouts. It was just one less stressor in my day. And no, I'm not the guy who doesn't like to interact with people; I just don't like small talk, or dealing with someone who's having a bad day. As a former Security Guard--I honestly didn't care about the property I was protecting. I would honestly help load the loot as long as I wasn't shot. I didn't protect any human life though. I saw a Microsoft built robot patrolling a car lot, or something of the like, on T.V., and realized a robot can be programmed "to just protect, and serve!". I could see how it could do a better job than a human, or me? I do see a computer running an established corporation in the future. It will be programmed to maximize stockholders returns, but take on risk. It won't marry. It won't need Therapy. It won't need to buy a yacht. As to computer programming, I look at Ruby on Rails. Just how easy will it be to put up a dynamic website in ten years? If Gates, Musk, and Hawkings are concerned--I'm concerned. AI dominance seems far off, unless you are a Golden Gate Toll Taker? I'm a nobody, but if I could add an admendment to the U.S. constitutuion it would be along the use, and limits of AI, and Automation.

[pjc50]

It's weird given that there are real civilization level threats: climate change, hostile unintelligent self-replicating human-hosted biological threats (diseases), nuclear war, and so on.

"How do we prevent AI destroying us?" is not as useful a question as "how do we prevent us destroying us?"

[TeMPOraL]

Unfriendly AI is one of such existential problems. Maybe further out in the future that the ones you mentioned, but with side effect of possibly taking the whole universe down with us if we get it wrong. And there will be people trying to pursue strong AI for various reasons, including help in fixing all the previous threats you mentioned.