LESSWRONG
LW

Comment Permalink

Doing an audit to catch all vulnerabilities is monstrously hard. But finding some vulnerabilities is a perfectly straightforward technical problem.

It happens routinely that people develop new and improved vulnerability detectors that can quickly find vulnerabilities in existing codebases. I would be unsurprised if better optimization engines in general lead to better vulnerability detectors.

See in context

16 [link] [poll] Future Progress in Artificial Intelligence

by Pablo

9th Jul 2014

1 min read

16

Vincent Müller and Nick Bostrom have just released a paper surveying the results of a poll of experts about future progress in artificial intelligence. The authors have also put up a companion site where visitors can take the poll and see the raw data. I just checked the site and so far only one individual has submitted a response. This provides an opportunity for testing the views of LW members against those of experts. So if you are willing to complete the questionnaire, please do so before reading the paper. (I have abstained from providing a link to the pdf to create a trivial inconvenience for those who cannot resist temptaion. Once you take the poll, you can easily find the paper by conducting a Google search with the keywords: bostrom muller future progress artificial intelligence.)

New to LessWrong?

16

[link] [poll] Future Progress in Artificial Intelligence

New Comment

89 comments, sorted by

top scoring

Click to highlight new comments since: Today at 8:32 PM

Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

[-]AlexMennen11y120

From the abstract of the paper:

Nearly one third of experts expect this development to be ‘bad’ or ‘extremely bad’ for humanity.

Where do they get this claim from? From the table in section 3.5 of the paper, it looks like they must have looked at the average probability that the experts gave for HLAI being bad or extremely bad (31%), but summarizing that as "nearly one third of experts expect ..." makes no sense. That phrasing suggests that there is a particular subset of the researchers surveyed, consisting of almost a third of them, that believes that the outcome would be bad or extremely bad. But you could get an average probability of 31% even if all the experts gave approximately the same probability distribution, and then there would be no way to pick out which third of them expect a bad result.

8danieldewey11y

Thanks, Alex; I think you're right, and am checking into it.

7danieldewey11y

...Vincent has now updated the paper; thanks again!

[-]Peter Wildeford11y90

I'm going to actually link to the paper, because it was actually non-trivially difficult for me to find, and because this page is now the top result for your suggested search query.

2Pablo11y

Sorry about that. The paper showed up first when I googled those keywords.

[-]Luke_A_Somers11y80

The most astonishing thing to me is what the paper gives as the responses to question 3, part B

"Assume for the purpose of this question that such HLMI will at some point exist. How likely do you then think it is that within (2 years / 30 years) thereafter there will be machine intelligence that greatly surpasses the performance of every human in most professions?"

EETN group, 30 years, median: 55%

What? They think that given a HLMI and 30 years, we have only a 55% chance make a SHLMI? Especially since they (on median) think it'll take 36 years to g... (read more)

3RomeoStevens11y

I feel like even experts reject anything "superhuman" based on stuff like absurdity heuristic or not thinking intelligence is a thing etc.

1[anonymous]11y

And now you've learned how little we can trust expert consensus on these issues.

0Daniel_Burfoot11y

I think you're misreading the question. It's not about going from human to superhuman intelligence, it's about exceeding human performance in various professions once human-equivalent intelligence is achieved. A lot of professions involve more than just intelligence. Almost all professions involve a large amount of human interaction, as well as the ability to navigate spaces and situations that are designed for humans.

1Luke_A_Somers11y

I am highly aware of the exact wording of the question, but thank you for checking. Your point makes it even more unbelieveable. Things do depend on which professions you see it as handling adequately in the starting case, keeping in mind that it has to be adequate at 'most'. Also, what you consider a 'profession'. Lawyer, doctor, scientist, engineer, programmer... manager, carpenter, plumber, mason, police, fire fighter... receptionist, trucker, waiter, cook, retail clerk... By the time you can be adequate at 'most' of those jobs, you've got the 'navigating human spaces' and 'interacting with humans' bit pretty well licked, so the difficulty of these matters is already handled by the time you start the clock. If on the other hand you cut it off after the second ellipsis (the last few are jobs but not 'professions') then you might be able to be only halfway decent at navigating spaces heavily optimized for humans, but you have to already be positively ace at interacting with people. If on the other hand you cut it off after the first ellipsis, and you're really just talking hard-core Professions with a capital P, then maaybe it can get away with being effectively immobile to begin with... but even in this extreme case, I can't see giving it the ability to walk, run, crawl, climb, shimmy, skoonch, and brachiate, or full equivalents, being a task that would span more than 30 years even taking the present as a starting point - and we haven't got HLMI today.

[-]gwern11y40

You could compare with the existing poll results from http://lesswrong.com/lw/jj0/2013_survey_results/

[-]James_Miller11y00

Starting in 20 to 30 years the most important AGI precursor technology will be genetic engineering or some other technology for increasing human intelligence. Any long term estimate of our ability to create AGI has to take into account the strong possibility that the people writing the software and designing the hardware will be much, much smarter than currently exist, possibly 30 standard deviations above the human mean in intelligence.

0V_V11y

So humans with IQ 550..., and 7.5 meters tall. I don't think that genetics work that way. Extraordinary claims require extraordinary evidence, not naive extrapolations from a binomial distribution model.

-2James_Miller11y

The 30 std link is to a blog of one of the world's leading experts on the genetics of intelligence.

2V_V11y

His Wikipedia page doesn't seem particularly impressive. Anyway, there are Nobel laureates who believe in homoeopathy and ESP, hence even if he was indeed a "leading expert on the genetics of intelligence", his word alone, especially a blog post, doesn't remotely look like enough evidence for such an extreme claim.

0James_Miller11y

The best evidence would be if animal breeders can move targeted traits by at least 30 standard deviations. If they do, the 30 std IQ claim wouldn't be extraordinary.

2gwern11y

Hsu has oft cited the example of oil content of plants. Another interesting example is breeding cows for milk: http://www.washingtonpost.com/blogs/wonkblog/wp/2014/02/17/cows-are-incredible-they-might-just-keep-producing-more-milk-forever/ The graph indicates that the average output per cow in 1990 (24 years ago) was ~15,000 pounds per year. It's unclear what the SD is, but given the peak mentioned of 50,000 pounds, an SD of 1000 pounds alone would imply breeding and other practices have shifted the cow population by an enormous amount over the very recent past. This itself seem to be a huge improvement over a lot of places: http://www.dairymoos.com/how-much-milk-do-cows-give/ gives a graph of unknown source http://www.dairymoos.com/wp-content/uploads/2013/08/image.png which compares cross-nationally; the USA is not at the top for average pounds of milk for cows, but what's interesting is contemporary Ethiopia has an entry indicating their cows are at 838 pounds. Ethiopia is a poor impoverished country with lots of traditional agriculture... so in other words, the normal human condition for most of history. If we set an SD as 838 (the Ethiopian SD is surely smaller, but let's go with it), then humanity has already improved cows by 17 SDs ((15000-838) / 838) and if we could get to the 50k some cows have performed at, then we would have improved milk yield by 58 SDs. There are probably a lot of comparable examples for horse speed, hog size, beef growing speed, etc, but I'm not an agriculture expert so this is just one I remember running into focusing on the incredible gains made in recent times.

1V_V11y

How much of that gain is genetic? If I understand correctly, cows in developed countries are feed nutrient-rich feed, such as soy and corn.

2gwern11y

I don't know, but if we can see easily >58SDs from a combination of carefully engineered environment and centuries of breeding... And I don't know how much of that is simply being fed nutrient-rich feed, given the doubling of productivity over the past 2 decades, unless dairy farmers only then realized 'oh, we should feed cows more!', which seems unlikely. That said, even just breeding is now old-fashioned; these days, the cutting edge in cow tech is using genotyping + phenotype data to more accurately estimate 'lifetime net merit' and pick animals to breed (a form of molecular breeding).

0gwern11y

Hsu links a review paper70305-4/fulltext "'Reliability of genomic predictions for North American Holstein bulls', VanRaden et al 2009") on the topic which discusses the usefulness of genetic prediction using cheap small SNP chips and how well it fits the highly-polygenic non-fixed model (which would imply that there's still a very long way to go before gains from breeding disappear):

0V_V11y

Thanks for the link. If I understand correctly, the inheritability of a trait often increases with a decrease of environmental variability. In this study they are comparing cattle raised in modern times in a developed country (the Netherlands, I think), hence the environment was likely about optimal, and unsurprising most of the observed phenotypic variation had a genetic origin. Ethiopian subsistence farmers probably don't have access to cheap soy and corn and have their cows graze on marginal lands, therefore nutrient availability is likely the limiting factor in their milk production. Similar patterns can be found in traits like human stature and IQ, which are more inheritable in developed countries rather than in third-world countries, and are subject to quick bursts when a country becomes more developed. Also, in the modern dairy industry, cows are slaughtered around the age four and sold for beef, while in subsistence farming they are likely to be kept for many years past peak milk production, resulting in lower lifetime averages. As for the specific of the inheritability of a continuous trait, I'm not an expert of genetics, but it seems to me that a polygenic model makes intuitive sense, as was quantitatively confirmed by this study. They found that a non-linear model predicts the data better than a linear model, which is however quite good, and again I don't find this particularly surprising since linear approximations often perform well on sufficiently smooth functions, especially in the neighbourhood of a stationary point (where you can expect the genotypes of a relatively stable population to be, approximately). My problem with Hsu line of argument is that he extrapolates predictions of these kinds of linear models way past observed phenotypes, which is something that has no theoretical basis, especially given that non-linear effects (logarithmic and logistic responses, square-cube effects, etc.) are ubiquitous in biology.

2gwern11y

Yes. (More relevantly, I'd say that as the environment gets better, the heritability will increase.) Overall, your points about the Ethiopian cows are correct but I don't think they would account for more than a relatively small chunk of the difference between the best American milk cows and regular Ethiopian milk cows. It really does look to me like humanity has pushed milk capacity dozens of standard deviations past where it would have been even centuries ago. Not surprising no, but people have seriously argued to me that things like embryo selection will not work well or at all because it's possible important stuff will be due to nonlinear genetic interactions (most recently on Google+, but I've seen it elsewhere). So it's something that apparently needs to be established. I'm not sure how seriously Hsu takes the 30SD part as translating to underlying intelligence; the issue of SDs/normal ordinal distribution of intelligence in the population vs a hypothetical underlying cardinal scale of intelligence (http://lesswrong.com/lw/kcs/what_resources_have_increasing_marginal_utility/b0qb) is not really easy to come down to a hard conclusion except to note that in some areas AI progress curves spend a while in the human range but often go steadily beyond (eg computer chess), which suggests to me that large difference in human intelligence rankings do translate to fairly meaningful (albeit not huge) absolute intelligence differences, in which case the 30SDs might translate to a lot of real intelligence and not some trivial-but-statistically-measurable improvements in how fast they can do crosswords or something. I think probably the best response here is to take it as saying that the lower limit will be extremely high and equivalent to the top observed phenotype, like a von Neumann. Since right now estimates of IVF sperm donor usage in the USA suggest something like 30-60k kids a year are born that way*, if the fertility doctors dropped in an iterated embryo selectio

2V_V11y

The most extreme example I'm aware of is the size of dogs: a chihuahua is about 4.3 times smaller than a gray wolf in terms of shoulder height, 4 stds of the wolf height. Increasing size seems much more difficult than decreasing it, as the tallest dogs, great danes, are in the same range of wolves. Some breeds of dogs are shorter and more massive than wolves, but not by much. Other domesticated species, AFAIK, show much less variance.

[-]XiXiDu11y00

I don't yet know how to update on this with respect to MIRI. One third of experts expect the development of human level AI to be ‘bad’. Well, I don't think I ever disagreed that the outcome could be bad. The problem is that risks associated with artificial intelligence are a very broad category. And MIRI's scenario of a paperclip maximizer is just one, in my opinion very unlikely, outcome (more below).

Some of the respondents also commented on the survey (see here). I basically agree with Bill Hibbard, who writes:

Without an energetic political movement t

... (read more)

4gwern11y

Correct me if I'm wrong, but wasn't the wording about what it was expected to me, not merely what was the worst-case scenario?

3XiXiDu11y

I have read the 22 pages yesterday and haven't seen anything about specific risks? Here is question 4: Question 3 was about takeoff speeds. So regarding MIRI, you could say that experts disagreed about one of the 5 theses (intelligence explosion), as only 10% thought a human level AI could reach a strongly superhuman level within 2 years. But what about the other theses? Even though 18% expected an extremely bad outcome, this doesn't mean that they expected it to happen for the same reasons that MIRI expects it to happen, or that they believe friendly AI research to be a viable strategy. Since I already believed that humans could cause an existential catastrophe by means of AI, but not for the reasons MIRI believes this to happen (very unlikely), this survey doesn't help me much in determining whether my stance towards MIRI is faulty.

8Luke_A_Somers11y

Hit the brakes on that line of reasoning! That's not what the question asked. It asked WILL it, not COULD it. If there is any sort of caution at all in development, it's going to take more than 2 years before that AI gets to see its own source code.

0[anonymous]11y

In many architectures the AI "sees" its source code as part of normal operations. It's a required fact for how these architectures are structured. Indeed, it is the only way the benefits of AI can be applied to AI itself, and the chief mechanism for an intelligence explosion. It's like having a nuclear program and saying "let's avoid using any fissionable materials." It could happen... but it would be rather missing the point.

1Luke_A_Somers11y

I'd dispute your last paragraph. An AI that doesn't examine and modify its own source code as its core architecture can still examine and modify its own source code if you let it. That's like keeping the fissionable materials wrapped in graphite like in a pebble-bed reactor. I would class baking it in from the start not as simply having fissionable material, but more as juggling bare, slightly subcritical blocks of Plutonium.

0[anonymous]11y

That's missing the point I'm afraid. What I meant was that the operation of the AI itself necessarily involves modifying its own "source code." The act of (generalized!) thinking itself is self-modifying. An artificial general intelligence is capible of solving any problem, including the problem of artificial general intelligence. And the architecture of most actual AGIs involve modifying internal behavior based on the output of thinking processes in such a way that is Turing complete. So even if you didn't explicitly program the machine to modify its own source code (although any efficient AGI would need to), it could learn or stumble upon a self-aware, self-modifying method of thinking. Even if it involves something as convoluted as using the memory database as a read/write store, and updating belief networks as gates in an emulated CPU.

0Luke_A_Somers11y

Having to start over from scratch would be a very significant impediment. Don't forget that we're talking about the pre-super-intelligence phase, here. So, no, I don't think I missed the point at all.

0[anonymous]11y

Gah, no, my point wasn't about starting over from scratch at all. It was that most AGI architectures include self-modification as a core and inseparable part of the architecture. For example, by running previously evolved thinking processes. You can't just say "we'll disable the self-modification for safety's sake" -- you'd be giving it a total lobotomy! I was then only making a side point that even if you designed an architecture that didn't self-modify -- unlikely for performance reasons -- it would still discover how to wire itself into self-modification eventually. So that doean't really solve the safety issue, alone.

0Luke_A_Somers11y

I was disagreeing that that architectural change would not be helpful on the safety issue.

-2TheAncientGeek11y

You put source code in scare quotes. Most AIs don't literally modify their source cor, they just adjust weighting .ir whatever...essentially data.

0[anonymous]11y

I don't disagree with this comment. The scare quotes is because the AI wouldn't literally be editing the C++ (or whatever) code directly, the sort of things that a reader might think of when I say "editing source code." Rather it will probably manipulate encodings of thinking processes in some sort of easy to analyze recombinant programming language, as well as adjust weighting vectors as you mention. There's a reason LISP, where code is data and data is code is the traditional or stereotypical language of artificial intelligence, although personally I think a more strongly typed concatenative language would be a better choice. Such a language is what the AI would use to represent its own thinking processes, and what it would manipulate to "edit its own source code."

0XiXiDu11y

If I have a statement "X will happen", and ask people to assign a probability to it, then if the probability is <=50% I believe it isn't too much to a stretch to paraphrase "X will happen with a probability <=50%" as "It could be that X will happen". Looking at the data of the survey, of 163 people who gave a probability estimate, only 15 people assigned a probability >50% to the possibility that there will be a superhuman intelligence that greatly surpasses the performance of humans within 2 years after the creation of a human level intelligence. That said, I didn't use the word "could" on purpose in my comment. It was just an unintentional inaccuracy. If you think that is a big deal, then I am sorry. I'll try to be more careful in future.

2Luke_A_Somers11y

The difference here is that you considered this position to strictly imply being against the possibility of intelligence explosion. One can consider intelligence explosion a real risk, and then take steps to prevent it, with the resulting estimate being low probability.

4ChrisHallquist11y

I should note that it's not obvious what the experts responding to this survey thought "greatly surpass" meant. If "do everything humans do, but at x2 speed" qualifies, you might expect AI to "greatly surpass" human abilities in 2 years even on a fairly unexciting Robin Hansonish scenario of brain emulation + continued hardware improvement at roughly current rates.

4gwern11y

Yes, that sounds like an expectation or average outcome: 'overall impact'. Not a worst-case scenario, which would involve different wording. I'm not sure how much they disagree. Fast takeoff / intelligence explosion has always seemed to me to be the most controversial premise, which the most people object to, which most consigned SIAI/MIRI to being viewed like cranks; to hear that 10% - of fairly general populations which aren't selected for Singulitarian or even transhumanist views - would endorse a takeoff as fast as 'within 2 years' is pretty surprising to me.

8Luke_A_Somers11y

I wouldn't be surprised if an AI went from human-level to significantly-smarter-than-human just by brute force of 'now that we've figured out what to do, let's do it HARDER', and this sort of improvement could be done in under 2 years. This is quite a different thing from a hard takeoff.

3XiXiDu11y

In the paper human-level AI was defined as follows: Given that definition it doesn't seem too surprising to me. I guess I have been less skeptical about this than you... What sounds crankish is not that a human level AI might reach a superhuman level within 2 years, but the following. In Yudkowsky's own words (emphasis mine): These kind of very extreme views are what I have a real problem with. And just to substantiate "extreme views", here is Luke Muehlhauser:

[-]gwern11y110

Given that definition it doesn't seem too surprising to me. I guess I have been less skeptical about this than you...

I don't think much of typical humans.

These kind of very extreme views are what I have a real problem with.

I see.

And just to substantiate "extreme views", here is Luke Muehlhauser:

It might be developed in a server cluster somewhere, but as soon as you plug a superhuman machine into the internet it will be everywhere moments later.

That's not extreme at all, and also not the same as the EY quote. Have you read any computer security papers? You can literally get people to run programs on their computer as root by offering them pennies! That's the sort of security environment we operate in. Every botnet with millions of computers is a proof of concept.

0XiXiDu11y

He wrote it will be moments later everywhere. Do you claim that it could take over the Internet within moments?

6gwern11y

With proper preparation, yes. To reuse my example: it doesn't take long to register an Amazon account, offer a high-paying HIT with a binary download which opens up a port on the computer, and within minutes multiple people across the world will have run your trojan (well-paying HITs go very fast & Turkers are geographically diverse, especially if the requester doesn't set requirements on country*); and then one can begin doing all sorts of other things like fuzzing, SMT solvers to automatically extract vulnerabilities from released patches, building botnets, writing flashworms, etc. At that point, the only real question is how stringently one wishes to define 'moments later' and 'everywhere'. (Within a few seconds? Hard to see how any plausible AI could copy its entire source code & memories over the existing Internet that fast unless it was for some reason already sitting on something like a gigabit link. Within a few minutes? More reasonable. Is there any real difference in safety? No.) * IIRC a lot of Turk HITs - like for psychology surveys - specify they want Turkers who are from eg America, to reduce variation or get more relevant answers, so Turk requires a declared country for each worker and lets requesters mandate being from particular countries. That said, there's a lot of incentive for Turkers to lie - to qualify for more and more high-paying HITs - so even if the AI were for some reason to restrict by country, it'd wind up with a lot of diverse foreign computers anyway.

0XiXiDu11y

Thanks. Looks like my perception is mainly based on my lack of expertise about security and the resulting inferential distance. Are there good reasons to assume that the first such AI won't be running on a state of the art supercomputer? Take the movie Avatar. The resources needed to render it were: 4,000 Hewlett-Packard servers with 35,000 processor cores with 104 terabytes of RAM and three petabytes of storage. I suppose that it would have been relatively hard to render it on illegally obtained storage and computational resources? Do we have any estimates on how quickly a superhuman AI's storage requirements would grow? CERN produces 30 petabytes of data per year. If an AI undergoing an intelligence explosion requires to store vast amounts of data then it will be much harder for it to copy itself. The uncertainties involved here still seem to be too big to claim that a superhuman intelligence will be everywhere moments after you connect it to the Internet.

2gwern11y

I don't think so. Consider botnets. How hard is it to buy time on a botnet? Not too hard, since they exist for the sake of selling their services, after all. Do they have the capacity? Botnets range in size from a few computers to extremes of 30 million computers; if they're desktops, then average RAM these days tends to be at least 4GB, dual core, and hard drive sizes are >=500GB, briefly looking at the cheapest desktops on Newegg. So to get those specs: 35k cores is 17.5k desktops, for 104tb of RAM you'd need a minimum of 104000 / 4 = 26k computers, and the 3pb would be 6k (3000000 / 500); botnets can't use 100% of host resources or their attrition will be even higher than usual, so double the numbers, and the minimum of the biggest is 52k. Well within the range of normal botnets (the WP list has 22 botnets which could've handled that load). And AFAIK CGI rendering is very parallel, so the botnet being high latency and highly distributed might not be as big an issue as it seems. How much would it cost? Because botnets are a market, it's been occasionally studied/reported on by the likes of Brian Krebs (google 'cost of renting a botnet'). For example, https://www.damballa.com/want-to-rent-an-80-120k-ddos-botnet/ says you could rent a 80-120k botnet for $200 a day, or a 12k botnet for $500 a month - so presumably 5 such botnets would cost a quite reasonable $2500 per month. (That's much cheaper than Amazon AWS, looks like. https://calculator.s3.amazonaws.com/index.html 17500 t2.medium instances would cost ~$666k a month.) I don't know. Humans get by adding only a few bits per second to long-term memory, Landauer estimated, but I'm not sure how well that maps onto an AI. It may not be able to move itself instantly, but given everything we know about botnets and computer security, it would be able to control a lot of remote computers quickly if prepared, and that opens up a lot of avenues one would rather not have to deal with. (Original AI connects to Internet,

0[anonymous]11y

Why does the AI have to transfer its source code? I assumed we were just talking about taking over machines as effectors.

0Lumifer11y

It will want to -- for safety, to avoid being vulnerable due to having a single point of failure. However the "take over in moments" refers to just control, I think. Getting the AI to become fully distributed and able to tolerate losing large chunks of hardware will take a fair amount of time and, likely, hardware upgrades.

[-]jimrandomh11y100

The two quotes you gave say two pretty different things. What Yudkowsky said about the time-scale of self improvement being weeks or hours, is controversial. FWIW, I think he's probably right, but I wouldn't be shocked if it turned out otherwise.

What Luke said was about what happens when an already-superhuman AI gets an Internet connection. This should not be controversial at all. This is merely claiming that a "superhuman machine" is capable of doing something that regular humans already do on a fairly routine basis. The opposite claim - that the AI will not spread to everywhere on the Internet - requires us to believe that there will be a significant shift away from the status quo in computer security. Which is certainly possible, but believing the status quo will hold isn't an extreme view.

1XiXiDu11y

My problem with Luke's quote was the "moments later" part.

4Baughn11y

I took that as hyperbole. If I were meant to take it literally, then yes, I'd object - but I have no trouble believing that a superintelligent AI would be out of there in a matter of hours to minutes, modulo bandwidth limits, which is 'instant' enough for my purposes. Humans suck at computer security.

0Luke_A_Somers11y

Yes, applying the SI definition of a moment as 1/2π seconds and the ANSI upper bound of a plural before you must change units, we can derive that he was either claiming world takeover in less than 10/(2π)^1/2 ≈ 3.9894 seconds, or speaking somewhat loosely. Hmmmmmmmmm.

0jimrandomh11y

Right, and I'm saying: the "moments later" part of what Luke said is not something that should be surprising or controversial, given the premises. It does not require any thinking that can't be done in advance, which means the only limiting input is bandwidth, which is both plentiful and becoming more plentiful every year.

-1XiXiDu11y

The premise was a superhuman intelligence? I don't see how it could create a large enough botnet, or find enough exploits, in order to be everywhere moments later. Sounds like magic to me (mind you, a complete layman). If I approximate "superintelligence" as NSA, then I don't see how the NSA could have a trojan everywhere moments after the POTUS asked them to take over the Internet. Now I could go further and imagine the POTUS asking the NSA to take it over within 10 years in order to account for the subjective speed with which a superintelligence might think. But I strongly doubt that such a speed could make up for the data the NSA already possess and which the superintelligence still needs to acquire. It also does not make up for the thousands of drones (humans in meatspace) that the NSA controls. And since the NSA can't take over the Internet within moments I believe it is very extreme to claim that a superintelligence can. Though it might be possible within days. I hope you don't see this as an attack. I honestly don't see how that could be possible.

7jimrandomh11y

This is not magic, I am not a layman, and your beliefs about computer security are wildly misinformed. Putting trojans on large fractions of the computers on the internet is currently within the reach of, and is actually done by, petty criminals acting alone. While this does involve a fair amount of thinking time, all of this thinking goes into advance preparation, which could be done while still in an AI-box or in advance of an order.

-1XiXiDu11y

Within moments? I don't take your word for this, sorry. The only possibility that comes to my mind is by somehow hacking the Windows update servers and then somehow forcefully install new "updates" without user permission. So if I uploaded you onto some alien computer, and you had a billion years of subjective time to think about it, then within moments after you got an "Internet" connection you could put a trojan on most computers of that alien society? How would you e.g. figure out zero day exploits of software that you don't even know exists?

4Lumifer11y

Well, what's going to slow it down? If you have a backdoor or an exploit, to take over a computer requires a few milliseconds for communications latency and a few milliseconds to run the code to execute the takeover. At this point the new zombie becomes a vector for further infection, you have exponential growth and BOOM!

4Nornagest11y

Wouldn't have to be Windows; any popular software package with live updates would do, like Acrobat or Java or any major antivirus package. Or you could find a vulnerability that allows arbitrary code execution in any popular push notification service; find one in Apache or a comparably popular Web service, then corrupt all the servers you can find; exploit one in a popular browser, if you can suborn something like Google or Amazon's front page... there's lots of stuff you could do. If you have hours instead of moments, phishing attacks and the like become practical, and things get even more fun. Well, presumably you're running in an environment that has some nontrivial fraction of that software floating around, or at least has access to repos with it. And there's always fuzzing.

2Lumifer11y

Also, nowadays if you can suborn the cell towers taking over all the smartphones becomes fast and easy.

-1jimrandomh11y

When you are a layman talking to experts, you should actually listen. Don't make us feel like we're wasting our time.

0[anonymous]11y

Care to address his valid response point in the 2nd paragraph?

0jimrandomh11y

Nornagest already answered it; the sets of software in and outside the box aren't disjoint.

5Lumifer11y

Given a backdoor or an appropriate zero-day exploit, I would estimate that it would take no longer than a few minutes to gain control over most of the computers connected to the 'net if you're not worried about detection. It's not hard. Random people routinely build large botnets without any superhuman abilities.

0[anonymous]11y

Most computers are not directly connected to the internet.

0Lumifer11y

Assuming we are not talking about computers in cars and factory control systems, this is a pretty meaningless statement. Yes, most computers sit behind routers, but then basically all computers on the 'net sit behind routers, no one is "directly" connected. Besides, routers are computers and can be taken over as well.

0[anonymous]11y

But not by that zero-day microsoft exploit you found. If your router is a cisco system, you need a cisco zero-day exploit to access the machines behind it, or some other way of bypassing the firewall. Sure, it could take over all the already vulnerable computers, the same ones which are already compromised by botnets. But I object to calling these "most of the computers connected to the 'net".

0Lumifer11y

The original scenario discussed was the NSA taking over the internet. I assume that the NSA has an extensive collection of backdoors and exploits (cf. Snowden) for Microsoft and Linux and Cisco, etc.

0[anonymous]11y

Yes well I thin XiXiDu did himself a disfavor there. If Snowden is to be believed and as various state-sponsored botnets (Stuxnet, Flame, BadBIOS(?)) have shown, the NSA has already "taken over" the internet. They may not have root access on any arbitrary internet-connected machine, but they could get it if they wanted. My objection (and his?) is against the claim that an AI could replicate this capability in "moments," according to the "because superhuman!" line of reasoning. I find that bogus.

4Lumifer11y

Let me suggest a way: * (1) Gain control of a single machine * (2) Decompile the OS code * (3) Run a security audit on the OS, find exploits Even easier if the OS is open-sourced.

2Nornagest11y

An AI probably wouldn't need to decompile anything -- given the kind of optimizations that one could apply, there's no particularly strong reason to think one would be any less comfortable in native machine code or, say, Java bytecode than in source. The only reason we are is that it's closer to natural language and we're bad at keeping track of a lot of disaggregated state.

0[anonymous]11y

That is a monumentally difficult undertaking, unfeasible with current hardware limitations, certainly impossible in the "moments" timescale.

4gwern11y

I think you underestimate the state of the art, such as the SAT/SMT-solver revolution in computer security. They automatically find exploits all the time, against OSes and libraries and APIs.

-2[anonymous]11y

I think you miss my point. These SAT solvers are extremely expensive, and don't scale well to large code bases. You can look to the literature to see the state of the art: using large clusters for long-running analysis on small code bases or isolated sections of a library. They do not and cannot with available resources scale up to large scale analysis of an entire OS or network stack ... if they did, we humans would have done that already. So to be clear, this UFAI breakout scenario is assuming the AI already has access to massive amounts of computing hardware, which it can re-purpose to long-duration HPC applications while escaping detection. And even if you find that realistic, I still wouldn't use the word "momentarily."

7jimrandomh11y

They have done that already. For example, this paper: "We implement our approach using a popular graph database and demonstrate its efﬁcacy by identifying 18 previously unknown vulnerabilities in the source code of the Linux kernel."

0gwern11y

Large clusters like... the ones that an AI would be running on? They don't have to scale although that may be possible given increases in computing power (you only need to find an exploit somewhere, not all exploits everywhere), and I am skeptical we humans would, in fact, 'have done that already'. That claim seems to prove way too much: are existing static code analysis tools applied everywhere? Are existing fuzzers applied everywhere?

2Lumifer11y

Why in the world would a security audit of a bunch of code be "monumentally difficult" for an AI..?

0[anonymous]11y

It requires an infeasible amount of computation for us humans to do. Why do you suppose it would be different for an AI?

0Lumifer11y

Um. Humans -- in real life -- do run security audits of software. It's nothing rare or unusual. Frequently these audits are assisted by automated tools (e.g. checking for buffer overruns, etc.). Again, this is happening right now in real life and there are no "infeasible amount of computation" problems.

0asr11y

Doing an audit to catch all vulnerabilities is monstrously hard. But finding some vulnerabilities is a perfectly straightforward technical problem. It happens routinely that people develop new and improved vulnerability detectors that can quickly find vulnerabilities in existing codebases. I would be unsurprised if better optimization engines in general lead to better vulnerability detectors.

9Adele_L11y

It's not like it's that hard to hack into servers and run your own computations on them through the internet. Assuming the superintelligence knows enough about the internet to design something like this beforehand (likely since it runs in a server cluster), it seems like the limiting factor here would be bandwidth. I imagine a highly intelligent human trapped in this sort of situation, with similar prior knowledge and resource access, could build a botnet in a few months. Working on it at full-capacity, non-stop, could bring this down to a few weeks, and it seems plausible to me that with increased intelligence and processing speed, it could build one in a few moments. And of course with access to its own source code, it would be trivial to have it run more copies of itself on the botnet. Even if you disagree with this line of reasoning, I don't think it's fair to paint it as "*very extreme".

0XiXiDu11y

With "very extreme" I was referring to the part where he claims that this will happen "moments later".

4Adele_L11y

Yes, that was clear. My point is that it isn't extreme under the mild assumption that the AI has prepared for such an event beforehand.

0[anonymous]11y

Really? It seemed really surprising to me that number was not higher. People are used to technology doubling in less than 2 years, and it is intuitively very straightforward that if you have a human-level AI running on 1,000 computers, than you could have a 1,000 * human-level AI running on 1,000,000 computers (not really because scaling relationships here might not be linear, but the linear assumption is a common intuition I expect most people to share), and two years is more than enough to build a bigger datacenter. There are two aspects of the Scary Idea which are controversial, and which I don't think this question covered: First, that an AI could inspect its own source code and take over the job of improving itself, thereby turning e^n improvement into e^(e^n) (something which has never happened before). This is generally accepted in the AGI community, but otherwise a foreign, non-intuitive idea. Second, that an AI could go from human-level to radically superhuman within days, hours, minutes, or even seconds. Few if any outside of MIRI believe this (and I can't get a straight answer as to whether they believe it either. If not, That Alien Message should be retracted.)

2gwern11y

People might expect there to be lots of AIs quickly, but not each individual AI to grow quickly. Remember, the typical case is that parallelization sucks hard and you get sublinear scaling after a lot of work which often tops out under a relatively small number of computers. That's why everyone was so unhappy about single-core performance version of Moore's law breaking down: we don't want to program parallelly. On top of that, a lot of people have intuitions about diminishing returns & computational complexity which suggest that throwing more computing power at an AI helps ever less. Is that generally accepted even just in the AGI community? That's another idea I usually see exclusively associated with Singulitarian communities. (As you say, it is controversial in general.)

0jimrandomh11y

Death of single-core Moore's Law has been greatly exaggerated. http://cpudb.stanford.edu/visualize/spec_int_2006

0[anonymous]11y

For most AGI architectures I've seen, the computationally expensive work is embarrassingly parallel. Programming solutions embarrassingly parallel problems is quite simple. I guess that depends on how "generally accepted" is to be interpreted. It is not as widely accepted as, say, plate tectonics is among geologists. It is certainly a view held among all OpenCog developers, including Goertzel. OpenCog itself is basically designed for recursive self-improvement. I also recall reading an interview with Hugo de Garis where he discussed a similar recursive self-improvement scenario. Hopefully someone can find a link. Talks on friendliness and hard-takeoff risk reduction are common at the AGI conferences. It's not a universal view however, as Pei Wang's NARS seems to be predicated on a One True Algorithm for general intelligence, which "obviously" wouldn't need improvement once found. Perhaps my view is biased towards the communities I frequent, as my own work is on how to turn OpenCog/CogPrime into a recursively self-improving implementation. So the people I interact with already buy into the recursive self-improvement argument. It is a very straight forward argument however: if you assume that greater-than-human intelligence is possible, and that human-level intelligence is capable of building such a thing, then it is straight forward induction that a human-level artificial computer scientist could also build such a thing, and that either by applying improvements to itself or staging it could do so at an accelerating speed. To such an extent that an AGI researcher accepts the two premises (uncontroversial, I think, albeit not universal), I predict with high probability that they also believe some sort of takeoff scenario is possible. There's a reason there is significant overlap between the AGI and Singulitarian communities. Where people differ greatly, I think, is in the limits of (software) self-improvement, the need for interaction with in the environment as part

2Shmi11y

Scott Adams half-jokingly suggests another way AI can lead to extinction through augmentation: This is by no means an original idea, of course, just not often listed as an AI risk.

4Lumifer11y

AI as the enabler of wireheading..?

2brazil8411y

Well that's already kind of happening, no? But a lot of people reproduce not because it feels good in the way that sex feels good but out of a longer-term sense of obligation. Well I think I would put it in the same category as wireheading. And given the choice, probably a lot of people will do it. Judging by the percentage of the population which declines reproduction in favor of sex with contraception; drugs; alcohol; internet porn; and other stimulation. But I predict some people won't make that choice, either because they like the idea of reproducing; or because their religion tells them to, or whatever. So I doubt it will lead to extinction, but it could very well re-shape humanity. Which seems to be happening already. If you do the math, ultra-religious types are on track to world demographic domination.

Moderation Log