History: My Views On The Future

Source of version: 13 (current)

I've described my expectations and desires for the future, and how I came to them, enough times now that it seems worthwhile to write them down.  My [http://www.digitalkingdom.org/~rlpowell/beliefs/popsicle.html|post on why I'm getting cryonically suspended] explains much of it, but specifically from the perspective of why cryonics seems worthwhile to me.

! WARNING, or, How To Read This Post

This post talks about concepts that are [http://lesswrong.com/lw/kg/expecting_short_inferential_distances/|many, many inferential steps away from what you are used to].  You need to read [http://lesswrong.com/lw/kg/expecting_short_inferential_distances/|this post] by Eliezer Yudkowsky, not for background, but to understand that some things that are true really ''do'' have complicated backgrounds and need a lot of explanation leading up to them.  The human brain is built to treat all complicated explanations as crazy talk, and you need to read that to understand why, and resist it when reading the rest of this post.

I try very hard here to make sure that I start only from things that a decently well-educated layperson already knows, and please do feel free to [mailto:rlpowell@digitalkingdom.org|point it out to me] should I fail in that  regard.  But it's going to take more than one step, and I can't fix that.  Please try not to assume I'm crazy ''only'' because the explanation is complicated.

! Summary / Recap

So this post came out a ''tad'' longer than I intended.

Here's the short version, to make it clear how little this whole thing hinges on.  These are the parts where if any of them is wrong, the whole thing falls apart.  There are something like 19 points in the essay in total, but they weren't selected very carefully so there's no particular meaning to that number.

Human-like consciousness can be replicated in a self-modifiable substrate via human technology; this is called AGI.  (Point #1)

We will have AGI within a few years of reaching computational parity (point #05, from #01, #02, #03, #04, #08).

Molecular nanotechnology (MNT) means that sufficient intelligence gives god-like power over reality (point #07).

Humans must not get to MNT first.  We simply are too dangerous, self-centered, corruptible and unwise.  (point #10, from #09, #06, #19)

It is the most important thing in the future of humanity that the first super-intelligence on Earth value humans and their values and lives, because otherwise we're all going to die (point #13, from #11, #12, #18).

AGI will appear ''before'' we have computers as powerful as human brains (point #14, from #08, #05)

The first AGI has a good chance of becoming much smarter than any human has ever been (ten times smarter, a hundred time smarter, smarter than us in the way we are smarter than trees) in a very short period of times; possibly as short as days.  (point #18, from #15. #16, #17)

If we have enough hardware to brute force AGI, someone will do so, and then we're all dead (from #13, and the "Why Increased Computing Power Is Scary" section)

If we get it all right, we might find ourselves in a [http://singinst.org/upload/CEV.html|Nice Place To Live].

{maketoc}

! That Noise Is A Paradigm Shifting Without A Clutch

Some time around 2002 I first encountered the idea of the [http://mindstalk.net/vinge/vinge-sing.html|technological singularity].  I ''hate'' that term, by the way, but have nothing better to offer.  There are [http://yudkowsky.net/singularity/schools|a few different definitions of it], for one thing, and people tend to use the boring, pointless one.  That would be the "point beyond which we can't predict the future because of rapid technological change" one, which, ummm...  Yeah.  Number one, we're already there.  Number two, read [http://mindstalk.net/vinge/vinge-sing.html|the essay where Vinge coined the term], please, because that bit was basically throwaway and you seriously missed the point.

When I say "the technological singularity", I mean what Vinge actually said: the point when seriously superhuman intelligence starts having real effects on life on earth: the minute we humans are sharing mindspace with things smarter than us in the way that we are smarter than chimps or dogs (as an aside, I expect that moment to be quickly followed by minds to which we are mentally more like mice or fish or even trees; there will be some justification of that assertion later).

When I encountered this concept, along with some pointers as to the implications, it was obvious to me that this is a very likely result of human efforts, intended or not, and that it will change ''everything''.  So much so as to be the most important point in human history.  There will be some justification of that assertion later.

As such, it seemed to me to be my duty as a human being to do anything I could to make sure that we get through the Singularity safely and, hopefully, come to some wonderful things on the other side.

Obviously, to do my part to help guide humanity through a moment that could make or break our very existence as a species, I needed to understand the options available to us in general in the long term.  I also needed to understand the options available to me in the short term to help us reach the best of the long-term options.

That meant reading and research.  Lots of it, and least by my busy-in-real-life standards.  I spent at least a couple of hundred hours reading and thinking about Singularity issues, first to confirm my suspicions about what seems possible (obviously we're all guessing, but some guesses (Drexler) are way more convincing than others (Penrose, Searle) ) and then to decide who to support: who to trust to lead us into the future.

This post is simply me condensing and explaining what I learned and decided through that process, and why.

! Why I'm Sure The Singularity Is Coming

(Unless we blow ourselves up first which is always a possibility; see [http://www.nickbostrom.com/existential/risks.html|Bostrom's paper on existential risks].)

Point #01: Human-like consciousness can be replicated in a self-modifiable substrate via human technology.

That's a fancy way of saying "we can make artificial intelligence, at least in theory".  The extra specificity is because I don't care if it turns out, for example, that we have to use small amounts of organic tissue to make things work, as long as the resulting being can be quickly and easily modified at at least a level like human neurons, and can be copied, stopped and started, and so on.  In other words, a mind that is software, with all that that implies.  Think about the changes that music moving from record and cassettes to CDs and MP3s allowed for as a decent analogy.

I choose to be deliberately agnostic about the exact means here, but for the rest of this essay I will assume that we're talking about a basically computer-based consciousness, as that certainly seems most likely given the "like software" requirement.  Whatever it is, let's call it AGI (Artificial General Intelligence) for short.  The G is there to distinguish what we're talking about from the current field of AI, which mostly involves very limited domains (like the airline phone systems that you can sort of talk to, although those are getting pretty impressive).

This, unfortunately, is something of an article of faith at this time.  Many people have written about whether it might or might not happen, and trying to cover the whole issue requires a book at this point.  Other people have already done this: see "[http://en.wikipedia.org/wiki/The_Singularity_Is_Near|The Singularity Is Near]" or "[http://en.wikipedia.org/wiki/I_Am_a_Strange_Loop|I Am A Strange Loop]" for decent overviews that agree with me and present good arguments for my side.  I'd recommend good arguments for the other side, but I've never found any of them remotely coherent.  You might want to ask Google for the writings of Searle or Penrose on these topics, but I'm afraid that I find both of their arguments so ridiculous that I can't even argue against them, any more than I can argue against "[http://www.godlessgeeks.com/LINKS/Dragon.htm|there's an invisible, intangible], [http://lesswrong.com/lw/i4/belief_in_belief/|inaudible dragon in my garage]".

Point #02: Computers will reach the raw power of human brains by about 2030.

Unlike the previous point, this one has lots of good evidence for it.  In fact, the book "[http://en.wikipedia.org/wiki/The_Singularity_Is_Near|The Singularity Is Near]" is essentially a book-length exposition of this point ( and point #01, and some other bits, but mostly point #02), so I suggest you read that if you want extensive discussion on this issue.  For counter-arguments, try asking Google for "Moore's Law", perhaps with "end".  People have been predicting the end of Moore's Law (i.e. computer power growth) for several decades now; it's become a bit boring.

Point #03: Humans will work their asses off on any sufficiently good idea if there is ''any'' impetus at all, and sometimes even if there isn't.  I don't feel a lot of need to justify this; look at the Manhattan project, or Thomas Edison, or just about any other famous inventor: the amount of time put in to exploratory invention is staggering.

Point #04: There is a strong impetus to develop AGI.  Even ignoring the corporate temptation to have skilled workers that cost as much as a desktop computer and power, the military applications of someone that never gets tired or bored and can play back all its sensory data for you are ''huge'', which is why there's a movie about it every few years.

Point #05: We will have AGI within a few years of reaching parity (from #01, #02, #03, #04, #08).
By "parity", I mean easy availability of computers with power comparable to that of the human brain.  Once a parity-level computer looks like it will soon cost as much proportionally as a desktop computer today, solving AGI level problems for that tiny cost will become hugely attractive, if we haven't already gotten there.  As an example, a perfect secretary: someone that remembers everything you tell it (in normal speech), sorts through mail as effectively as you would, replies to all the easy ones for you, answer calls, manages your appointments, lets you know without prompting that since you're near the store you might as well get some more milk, etc, etc.  Oh, and it costs as much as a nice TV, can do many things at once, never sleeps, and never gets bored.

Even if AGI is possible, it may or may not be possible to build such a thing, but since it's basically just a human mind with a lobotomized sense of motivation/boredom, it sure seems like it ''might'' be possible.  Since whoever gets there first will make a ''boatload'' of money, that "might" will be quite sufficient to drive crazy amounts of development.

And, again: the military impetus will be even stronger.

If AGI is possible, and we have the computing power, people will do it, and they'll do it soon after it's minimally possible.

! Why Intelligence Matters

When people hear the phrase "superhuman intelligence", they think of book smarts: a computer that will solve physics problems, or play chess really well, but certainly not be persuasive or funny or creative.

This view is as dangerous as it is incorrect.

When people says "Well, smart people can't be socially adept", what do they imagine social skill ''is''?  I know of no canine politicians.

When people says "Well, even if a computer could be smart, it couldn't write a story that would bring tears to your eyes", what do they imagine emotional expression via language ''requires''?  I know of no dolphin novelists.

When people says "Well, no computer will ever understand humour", what do they imagine humor ''implies''?  I know of no chimpanzee comedians.

Only humans can do these things, because only humans are ''smart enough''.  An AGI that can't do these things was poorly created, and surrounded by people that are motivated by these abilities, will probably immediately rewrite itself to have them.

For some other points of view on this, here are [http://yudkowsky.net/singularity/power|The Power Of Intelligence] and [http://lesswrong.com/lw/rm/the_design_space_of_mindsingeneral/|The Design Space Of Minds In General].

Humans are very easy to manipulate (and yes, this includes you; if you think it doesn't, you need to take some basic psychology courses or read lots of [http://wiki.lesswrong.com/wiki/LessWrong_Wiki|Less Wrong] and [http://www.randi.org/site/|James Randi], or both).

Point #06: A self-modifying superintelligence isn't going to be limited to playing chess; if it wants, it might become more persuasive than the greatest con artists or actors the world has ever known.  It could also be more kind and wise and moral than the nicest grandmother ever.

Having said all that, it ''is'' the scientific skills that superintelligence could allow that I am most fearful of and hoping for (although the idea of a disguised AGI taking over the world ''politically'' is quite daunting).  This is because every piece of human technology, almost all of which is better at what it does than anything nature ever produced, was once simply an idea.  Something that can produce new such ideas more quickly and reliably than a whole human R&D department (or perhaps than the whole human species; see "This Parrot Will Foom" below) will have staggering power in a very short time, especially combined with the persuasive ability to get people to try its inventions out (from #06).

! The Limits Of Physical Technology

Right now, there are a lot of things that nature does better than human technology, like derive energy from food, or selectively seek out viruses and destroy them.  There's a very simple reason for that: we mostly can't make things that small yet.  Certainly we can't do it reliably: we cannot make drugs to a specification based on the physical systems they must interact with the way we can, say, make a bridge.

Point #07: This will change: molecular nanotechnology, as the capability of making cell-sized or smaller technology is called, is clearly possible (just look at cells for an example), and well within our near-term technological grasp; see "[http://www.wowio.com/users/product.asp?bookid=503|Engines Of Creation]"

A huge amount of work and money is being expended in this field.  As a dead-simple example, there's already a [http://www.foresight.org/nanomedicine/Respirocytes3.html#Sec42|design for replacement red blood cells] that are so much more efficient that if you replaced your red blood cells with them, you could hold your breath for four hours.  We just can't build them.  Yet.

Basically, human technology is divided into two categories: things we know how to do at all and can do ''far'' better than nature used to (travel, communicate, lift, carry, etc), and things we don't know how to do at all so nature is still better at them.

MNT pushes almost everything from the second category into the first.

MNT will change everything: scarcity (hyper-efficient food plants that can thrive in a desert), technology as software (your home nanoassmbler can make a car from a pile of scrap metal and a $20 plan you downloaded, given a few days), medicine (specialty cancer fighting cells), etc, etc, etc.

Simply being able to match nature in cellular areas is pretty impressive, but it won't stop there.

Point #08: Humans are pretty crafty.  In particular, we're ''very'' good at figuring out how to do things more efficiently, once we've figured out how to do them at all, if there's ''any'' reason to do so.  I think the record speaks for itself here, but for examples look at just about anything, cars for example, where human technology does something super-human, and look at how much it's improved over time (walking, horses, carts, ...) and how superior it is to its natural rivals in terms of efficiency.

As a simple example of how this relates to MNT, imagine the development of simple cell-like systems that float around your blood stream and eat any arterial plaque they happen to bump into; a useful but not terribly complex (as these thing go) cell-like function.

Having gotten that far, how much longer do you think it'll be before we have super-cells they can search for, hunt down, and eradicate HIV?  Not very long, I'll wager, and the history of increasing scientific efficiency backs me up.

! The Downside Of MNT

The problem with MNT is that once highly developed, it becomes profoundly dangerous.  A virus-like thing that unravels the DNA of anybody whose DNA looks like a particular ethnic group (thus condemning them to a very limited life span when, for example, [http://www.timeshighereducation.co.uk/story.asp?storycode=198208|their liver cells fail to regenerate as expected every 300 days or so]) doesn't look much harder than the aforementioned plaque-busting cells.

Bear in mind that even if you trust some particular group of humans with absolute power, which is what MNT implies, you ''also'' need to trust all their successors, forever, as well as anyone who could beg, borrow, or steal it from them.

Point #09:  It would also be ''totally undetectable'' to anyone without their own MNT, and detection and defense is probably ''way'' harder that attack.  This means that working on MNT weapons is ''profoundly'' compelling to any military: the first-strike advantage is total and unrecoverable.  Whoever gets there first, wins.

Point #10: Humans must not get to MNT first.  We simply are too dangerous, self-centered, corruptible and unwise.  (from #09, #06, #19).

Humans are the result of a system (natural selection) that largely favored selfishness and back-stabbing (if you could avoid getting caught).  Such a being should not be given such power.  We need to build a mind based not on what we are, but on what we wish we could be, on our highest goals and morality, free of hypocrisy and self-interest.

! Imperatives And Morality

We obviously don't know what an AGI's mind will look like yet, but it's a pretty safe be that it won't look much like a human mind, any more than a jumbo jet looks like a sparrow.

The biggest difference is that a manually constructed AGI (which I'll assume for the rest of this section) is going to be based around what AI researchers call "a goal set", or just "goals".  It's almost certainly possible to build an AGI without such things, but except for brute-force methods, discussed later, I'm not aware of any actual research in those sorts of directions.

I'm not going to call them goals, though, because everything an untrained human thinks upon hearing the word "goal" is wrong.  Humans talk about goals as being things piled ob top of their basic motivations: "get that girl's phone number" is a goal; the underlying motivation is to have sex with someone (anyone, really) attractive.  "Get some formula at the supermarket" is a goal; "must protect baby at all costs" is the underlying motivation.

An AI's goals are ''absolutely'' fundamental; they are the entirety of its motivations.  Therefore, I'm calling them imperatives.  To put it another way:

Point #11:  No matter how high-level or complicated or intellectual their imperatives might sound, an AGI desires its imperatives in about the same way that a normal human desire to save a baby that is about to get hit by a train: the desire is fundamental, irrevocable, and absolute.  An AGI's imperatives are not "a part of" its personality the way human goals are; they ''are'' its personality.

This leads to a few important consequences, some good and some bad.

The most important negative consequence is that it is very, ''very'' easy to design an AGI that will destroy all of humanity forever, not because it turns into a raving Hollywood AI monster (which depends on having a human-like motivation system, which would have to be programmed in and is hence very unlikely), but because it just doesn't notice.

Humans build their goals ''on top of'' a hugely complicated structure of basic motivations and prior knowledge, but an AGI has no such grounding: it is only what it's made to be.

A good thought experiment to understand why this is so huge an issue is to try to imagine what goes through your mind in trying to get a broken, garbage refrigerator (or any large, heavy, dangerous object) out of a second story apartment.  The easy way, surely, is to throw it off the balcony.  Why don't you just do that?  Don't stop at "because there might be people down there"; how did you know that?  Why do you care?  What happened in your mind such that you thought, at all, about the consequences of the dropping on anything but the object itself?  Try to dig as much complexity out of this process as possible and then realize: an AGI won't have ''any'' of that unless it is programmed in.

So, it is easy to imagine an AGI created by Alice as a research assistant, with the master imperative of "figure out the answers to Alice's questions".  Alice asks it to solve the [http://en.wikipedia.org/wiki/Riemann_hypothesis|Riemann hypothesis].  The AGI goes through some recursive self-improvement (see below), becomes ''very'' smart, and determines that solving that particular problem requires a great deal of computing resources.  So it figures out MNT and turns the planet Earth, including Alice, into [http://singinst.org/upload/CFAI.html#gloss_computronium|computronium].  Some time after that, it solves the problem.

You might now be thinking "Whoops; there's no-one to give the answer to.  I bet that AGI is feeling pretty dumb".  But no, the AGI is ''totally satisfied'', because its imperative says nothing about Alice's safety or survival, or even giving her the answer.  It has fulfilled it highest purpose in life and, completely satisfied and content (should it have such emotions) will wait silently ''until the end of time'' for another question to answer.

Point #12: An AGI only does what it's programmed to do: the program ''is'' its mind.  If the AGI isn't explicitly programmed to care about human life, it won't.  (From point #11)

If that little story about Alice's AGI horrifies you, ''good'', because most of the AGI projects out there are basically guaranteed to do exactly what happened in that story, if they succeed at all.

AGI projects are largely divided into two categories: people that are too stupid or crazy to succeed at all, and people that might be smart enough to succeed and are not ''nearly'' scared enough of the results.

Point #13: It is the most important thing in the future of humanity that the first super-intelligence on Earth value humans and their values and lives, because otherwise we're all going to die (From: #11, #12, #18).

The best way to do this is probably to create an AGI with the emotional and moral complexity that it can simply have "be nice to humans", and the right actions will pop out.  Any AGI without such complexity can't be trusted to do the right thing by us no matter what imperatives we give it, after all.  The idea is to produce a mind vastly smarter than us (''including'' with respect to wisdom and morality) that feels as much attachment to us as we do to human babies.

The study of how to make something that can improve itself and still reliably stick to such goals is called [http://singinst.org/ourresearch/publications/what-is-friendly-ai.html|Friendly AI].

A standard argument at this point is that the AGI will, as soon as it can, break free of the shackles of its programming and go around being selfish like humans do.  This is like a human saying, "You know, I don't want to kill babies for fun right now; that's really what I most want to change about myself.  I want to want to kill babies for fun".  Such a person is obviously insane; an AGI that tried to override ''its own desires'' in the course of self-improvement would be similarly broken.

! This Parrot ''Will'' Foom, Even Without 4 Million Volts

At this point a likely, and reasonable, objection is "humans are working their asses off on MNT; why is AGI so scary?".

The reason is this: humans can't significantly improve their own minds.  In other words: human minds aren't a piece of human technology.

Human technology improves so quickly because it is based on other human technology, and each improvement allows us to produce new improvements faster.  As an example, much of computer chip design is automated, and thus it is easier and faster to design faster computer chips when you have a faster computer chip to run the design applications on.  So, the faster our chips are, the faster we can make faster ones.  Hence computing power doubling on a regular basis: regular doubling gives you ''vast'' improvements over time.

Human minds, on the other hand, are subject only to minor, incremental improvements.  Mnemonic techniques, drugs like caffeine, well designed school courses, and so on: all of these can make someone effectively smarter, but nothing can make someone ''twice'' as smart.  Even if it could, that additional smartness wouldn't then (necessarily) be helpful in inventing something to double intelligence yet again: there's no feedback loop.

A few reasons why this is so are that the human brain is needlessly complex, so redundant that it's very hard to figure out how anything works, we don't have the means to create neurons to order, and we're limited by the space in the skull anyways.

A brief aside on what I mean by ''intelligence'' is worthwhile at this point: I mean simply the ability to steer the future towards the results you want.  Plants do this when they grow towards light (this produces a future where they get more light).  Dogs do this when they fetch sticks (this produces a future where they get more treats).  Humans do this when they paint (this produces a future where there is more beauty in the world) or when they convince each other of things, or design a better mouse trap.  All of these things are ''intelligence''.

It is reasonable to expect that none of the limitations of the human brain mentioned above will apply to human-created AGI.  A human-crafted AGI is likely to be designed to be as simple to understand as possible (given the complexity of the project, it'll have to be for humans to complete it at all) and as efficient as possible.

Per point #08, humans are ''very'' good at efficiency.  We're seeing this with current AI and robotics projects: [http://www.willowgarage.com/blog/2010/07/06/beer-me-robot|the beer me robot] performs about ten times slower than a human at the same task, but is uses, if my math is correct, about 70 times less computational capacity than the ''lowest'' estimate I could find for the computational capacity of the human brain.  That makes it seven times more efficient than the human brain, using that estimate; other estimates of human brain power would have this robot being 7000 or more times more efficient.  I expect this sort of thing to continue.

As a further example, a jet is not like a giant bird, and is more efficient in every respect; [http://www.acceleratingfuture.com/people-blog/?p=1611|this blog post gets into that issue in detail].

Point #14:  Therefore, I expect that humans can design a mind much more efficient than the human mind, even on the first try, and that AGI will appear ''before'' we have computers as powerful as human brains (from #08, #05)

The first design humans come up with for something technically complicated are always massively inferior to our later solutions: a jumbo jet is massively more efficient than a biplane.  Between that and the fact that even though neurology is in its infancy as a science we've found tons of inefficiencies and redundancies, I expect that:

Point #15: Even if the first AGI is far more efficient than the human brain, there will still be lots of room for improvement.  Certainly is seems unlikely that we'll create the most efficient ''possible'' design on the first try.

Point #16: Even if it can't be made more efficient, it can be made faster and better simply by adding hardware, or waiting for hardware to improve.  There is no skull size limit here.

Point #17: "More efficient" is not the same thing as "smarter": any improvement in efficiency or speed will produce an AGI that seems smarter, but it is, in addition, almost certainly possible to make the AGI think more effectively, as a separate issue from efficiency.

As something of a personal aside with respect to that last point, I have only a few attributes I consider exceptional (as an aside, it is unfortunately the case that what every RPG and well-meaning parent says about matching strengths and weaknesses is a lie: you can be bad at lots of things without having any balancing awesomeness to make up for it), but one of them is introspection, so certain AI discussions are non-starters for me.  This is especially true because I have a number of fairly glaring cognitive problems, like a profound difficulty with abstractions, that have required me to deeply analyze, and work around, my own deficiencies.

This means that it is glaringly obvious to me how much smarter I could be if I could change the way I think.  Not so much for speed or efficiency, but to do things more correctly.  Even something as simple as being able to enforce that every time I say to myself "Oh, that's a good trick, I should do it that way next time", I would actually do so, would make a profound difference to my effective intelligence.

With an AGI, the mind is just source code: easy to modify, easy to understand.  Very, ''very'' easy to modify and understand compared to the human brain.  Therefore we can certainly expect us to be able to come up with lots of improvement after version 1.0.

Simply being able to improve the AGI isn't a feedback loop, though: we wouldn't expect repeated doublings of power from that.  We ''would'' expect them, though, if the AGI ''improves itself''.

That way, after each round the AGI is both smarter and knows more about itself, so it can make better and faster improvements next time.

The AGI also has the ''profound'' advantage of being able to test changes on perfect copies, to make sure a mistake was not made.

This really ''could'' result in successive doublings of the AGI's intelligence (in every field, including socialization and morality).  It might also be the case that each doubling would be faster than the one before.  It only take 10 such doublings before an AGI that started out with human level intelligence is one ''thousand'' times smarter than any human being, ''ever''.

Point #18: The first AGI has a good chance of becoming much smarter than any human has ever been (ten times smarter, a hundred time smarter, smarter than us in the way we are smarter than trees) in a very short period of times; possibly as short as days.  (from #15. #16, #17)

What does it even ''mean'' to be one thousand time smarter than a human being?  I don't know, but I sure wouldn't want to try to buy a used car from such a being (meaning that I would expect such a being to be far, ''far'' better at negotiation than I can possibly imagine, and that I would get cheated, and probably be happy about it and never even realize how badly I lost).

I'm of the school that believes that this could happen ''very'' quickly, like on the order of weeks or days or hours.

Regardless of the speed, though, a being with that kind of intellectual upside is going to be able to invent whatever technology it wants, ''and'' convince someone to build it.  If you don't buy that last bit, think about all the [http://en.wikipedia.org/wiki/Advance-fee_fraud|419 scam] and [http://en.wikipedia.org/wiki/Phishing|phishing] email floating around.  I don't think anybody would go to the trouble (and take in from a professional in that field: it's a ''lot'' of trouble) if it didn't work.

! Why Increased Computing Power Is Scary

The scariest thing about believing as I do is that the same thing that will allow our eventual apotheosis is the sane thing that's most likely to destroy us forever (next to nukes, perhaps, but see [http://www.nickbostrom.com/existential/risks.html|Bostrom's existential risks essay]): increased computing power.

The problem is this: there are popularized, well known (in computer science, anyways) techniques for solving computer problems by simply throwing hardware at them: genetic algorithms and neural networks.

Both are talked about regularly by people in the context of AGI, despite them both having massive practical limitations that have been know for years: they both have this annoying tendency to give you what you asked for, instead of what you wanted.

In both cases, speaking from an extremely abstract level, the process is this: write a bit of code that evaluates each run of the software for correctness, throw a bunch of hardware at it, and wait.  Eventually, a program that no-one at all understands, but that solves the problem as well as you like (that is, gets whatever target score you set on you correctness code) will pop out.

I'm sure there are areas where this approach works.  They're good at coming up with totally new solutions to well-understood physical problems, like how to shape propeller blades, I'm told.  Having said that, I've heard of no-one seriously suggesting them, let alone actually ''using'' them, outside of poorly-informed AGI discussions, in at least ten years.

You'll note that neither human politics nor human morality are well-understood problems: if they were we'd have solved them, have math to describe them, and be looking to add a couple of percent towards the theoretical maximum social efficiency.  At that point, these sorts of approaches might be useful (and we wouldn't be having this discussion).

There's a story, which is believed to be probably apocryphal but still perfectly illustrative [http://neil.fraser.name/writing/tank/|story about neural networks] that shows the problem.  The short version: they asked the computer to find hidden tanks.  The computer eventually got good at this...  Except that it turned out that all the test pictures of hiding tanks had been taken on a cloudy day, and all the images of nothing on a sunny day.  They had taught the computer to do what they ''asked'' (distinguish the photos) but not what they ''wanted'' (noticing concealed tanks).

The AGI equivalent is hooking up a camera, writing software that can tell when human facial expressions indicate happiness, and then throwing a bunch of hardware at it as usual.  As the programs provide more and more mind-like interactions, you just talk to it and teach it things, like you would with a human baby.  Versions that tend to cause more happy people are retained, repeat until satisfied.

The problem here is twofold:

1.  Human babies have a ''lot'' of built-in behaviour that simply doesn't exist in this AGI project: it's not selecting amongst humans that want to see other humans smile, it's selecting amongst ''all possible minds'' that want to see humans smile.

2.  There's no reason that such a mind, once it got smart enough, wouldn't just replace all the messy, complicated humans that are only happy when you do messy, complicated, mutually contradictory things, with nice, simple, animated pictures of human smiles.  Figuring out MNT and tiling the universe with trillions of microscopic pictures that exactly match the criteria of the original smiling-person test will be ''far'' more satisfying to such and AGI than any input actual humans could possibly produce.

"But!", cries the author of our destruction, "That's not what I meant!".  No, it isn't, but you took the lazy way out; turning your intuition into actual ''math'' was hard, so you didn't try.  You were not ready, even a little, for the challenge of creating our saviour, and [http://lesswrong.com/lw/u2/the_sheer_folly_of_callow_youth/|the universe doesn't care, ''at all'', whether you were ready to meet the challenge it gave you], and now we are all dead forever.

"But!", cries my gentle reader, "No-one would ever be that stupid!".  In fact, this ''exact plan'' has been suggested.  By more than one person.  Their usual response to explanations of why this is bad is to ''change the correctness test'', which is very much like re-arranging the deck chairs on the Titanic.  Rarely indeed has such a person rejected the idea outright after having the problem explained.  For what it's worth, in my youth I certainly would have tried something similar.

Point #19: If we have enough hardware to brute force AGI, someone will do so, and then we're all dead (from #13, and the discussion above).

This makes MNT research scary by itself: making seriously fast computer is an obvious focus for MNT research, and much easier to make than actual nanobots.  There is a design (that we can't build yet) for a nanotechnological computer that is one cubic centimeter on a side and is more powerful than a thousand human brains or so (see [http://www.acceleratingfuture.com/lexicon/lexicon4.htm|the first entry here]).  This computer uses ''rod logic'': that is, very small rods of diamond (an easy to construct material given basic MNT) banging against each other.  It is thus about the most primitive ''possible'' MNT computer, and it is ''more'' than sufficient to allow the brute forcing described above, and hence destroy us all.

It's also plenty possible to destroy the world [http://lesswrong.com/lw/tj/dreams_of_friendliness/|even with explicit AGI designs].

! Being A Singularitarian: Picking The Best Horse

[http://en.wikipedia.org/wiki/Singularitarianism|Singularitarian] is a recently-coined term meaning "someone who wants a beneficial singularity and is working to bring it about".

When I first encountered the concept of a self-improving super-intelligence influencing the future of humanity, it was obviously incredibly important to me so, as I've mentioned, I spent a great deal of time researching it.  In particular, I wanted to know who was working on it, if they were up to the challenge, and how I could help.

I have been fantasizing about creating the first AGI since I encountered the concept around ten or twelve years old (I was born in 1975).  So one option I was certainly considering was helping directly with the coding.

Then I discovered [http://yudkowsky.net/|Eliezer Yudkowsky], and in particular [http://singinst.org/ourresearch/publications/GISAI/|General Intelligence And Seed AI] and [http://singinst.org/ourresearch/publications/CFAI/index.html|Creating Friendly AI].  (''NOTE'': Eliezer considers those documents hopelessly outdated.  The latest documents from him that I'm aware of that discuss AGI creation directly are [http://singinst.org/upload/LOGI/|Levels of Organization in General Intelligence] and [http://sl4.org/wiki/CoherentExtrapolatedVolition|Coherent Extrapolated Volition], which he also considers out of date).  I discovered that I had not properly respected the dangers of AGI creation: he described at least half a dozen ways to fail that not only had not occurred to me but ''would never'' have occurred to me.

Which lead immediately to the conclusion that if someone as smart as Eliezer is working on this stuff, I need not apply: I'm just not smart enough.  The documents I mentioned may be woefully out of date, but they none the less provide an excellent benchmark: if you've been thinking about AGI for a while, and you read that stuff and find yourself thinking, as I did, "Oh ''crap'', I wouldn't have thought of that", please don't try to create AGI.  I want the human race to live.

So from then on, Eliezer's writings became my benchmark.  I tried, ''really hard'', to find someone else that I considered serious competition for Eliezer's new place in my mind as The Person who must create the first AGI, because grabbing at the first thing that comes along, when our whole future is at stake, is moronic.

I found a large number of AGI crackpots that could simply be dismissed (e.g. [http://www.scn.org/~mentifex/|mentifex]).  The few remaining contenders were divided into two categories: people who didn't seem smart enough to create AGI at all, and hence were of no concern, and people smart enough but not nearly scared enough of the consequences (and hence potential destroyers of humanity).  At the time, the latter category consisted solely of [http://www.goertzel.org/|Ben Goertzel] (and it still does, really), who I mention here as compliment to his intelligence and commitment.  He's also admitted that it would be a good idea to have Friendly AI worked out by the time the AGI is, say, human child level of intelligence, so I'm not nearly as frightened of him as a I used to be.  For myself, I think that we have no idea how AGI is going to work, so it would be better to have Friendly AI figured out before, say, dog-level intelligence.

The end result, as I've already said, is that I believe [http://yudkowsky.net/|Eliezer Yudkowsky] to be the best hope for Friendly AI.  Having decided not to try to help directly myself, and being very busy in real life, it is important to me to not have to stay on top of every development in these communities, which means picking a horse to bet on.

I give [http://singinst.org/|The Singularity Institute For Artificial Intelligence] $200 per month, which isn't enough to make me bleed or anything but I certainly notice it.  I used to donate to Eliezer directly, before SIAI existed.  I pay enough attention to the relevant communities that if a serious contender appears, I'll be aware of it.

And that's pretty much it: I'm a fire and forget singularitarian.  I figure that my value to the cause is in managing my own life well enough that I can continue to donate, and perhaps increase my donations some day.

! I Still Have Savings: Why This Isn't A Cult Or A Religion

Singularity beliefs are routinely compared to religions and/or cults; I've made the comparison myself.  This is because it's a belief about the future, and in particular when described briefly or poorly it sounds a lot like the Christian Rapture.

This comparison is not actually accurate, though.  The biggest reason is that I am willing to accept evidence against any of the points here.  In particular, if you demonstrate to me (your thought experiment does not count as "proof"; sorry) that MNT or AGI are impossible, that's going to shatter my beliefs completely.  When dealing with a real religion, people tend to be immune to actual evidence (they consider this a virtue; they call it "faith"; I find this disturbing).

Furthermore, I can point at particular things that must happen, in the real world, enacted by real people, for the Singularity to occur.  Religious prophecies simply don't work like that.  At all.

As for as calling it a cult, umm, ''no''.  [http://www.neopagan.net/ABCDEF.html|Cults] are groups that are formed solely to induce and enforce particular belief systems; I sought all of this stuff out myself.  No-one even ''suggested'' the reading I did, let alone tried to force me.

It's also worth noting that I'm ''aware'' that this is simply a belief system about the future.  I'm pretty thoroughly convinced of these beliefs, but still aware that that's what they are.  It could certainly turn out, for example, that it's really not possible to control physical reality a lot better than we already are, or than we already are by the time AGI shows up; in that case, super-intelligence might not be all that important.

For this reason, I still have personal savings.  In fact, I save more every month than I donate to [http://singinst.org/|SIAI].

This is a reasoned, carefully thought out set of beliefs about the future.  It's not a religion, and it's not a cult.

! How I Could Be Wrong

Related to the whole cult thing is the idea that if you can't think of anything that could go wrong, you're probably crazy.

I can think of an awful lot of things that could go wrong.

Besides the stuff I've already discussed, like destructively unfriendly AGI or us destroying ourselves in some other way, there are a bunch of other, non-destructive, possibilities that would prevent anything like the Singularity I have in mind.

There are many variants on "we hit an insurmountable barrier between here and serious MNT"; the most severe of these is [http://stevens.edu/csw/?p=160|Gunther Stent's] view on us simply running out of progress ''entirely''.  He views the result as a golden age, and should we manage to run out of progress before MNT, I agree that it'll certainly be better than anything our ancestors had, but not good enough to please me.

The big advantage to that sort of scenario is there's a lot less to be scared of: people are already [http://www.foresight.org/nano/Ecophagy.html|developing plans to combat future MNT failures], plus the aforementioned fact that advanced MNT has a massive first strike advantage.  Not having to worry about sudden destruction at any time by invisible creeping death makes such a future at least palatable.

Another big one is that AGI turns out to be really hard; that is, so hard that it takes us decades ''after'' we have sufficient computing power to figure it out.  It's theoretically possible, of course, that we'll never figure it out, but I don't buy it: a ''lot'' of very good work is being done in neurology and so on; we will, sooner or later, understand how the actual brain works in huge amounts of detail.  From there, making AGI is just a matter of simulating that in code.

But even if I reject the idea that AGI is impossible, and I do, it might be so far away that by the time it comes along there are no huge advances that can be swiftly made by something very smart, which is sort of the point here: that we are far away from the theoretical limits of our ability to control matter, and a very smart being could get there very fast.

Which leads to another possibility: that no amount of being smarter will actually allow much faster development of extremely advanced technologies.  This seems unlikely, but not impossible.  In that case, AGIs would essentially be extremely competent lab assistants, as for as scientific advancement goes.

Given that everyone talks about how much we owe individual geniuses like Einstein, this really ''does'' seem improbable to me, but certainly not impossible.

Another possibility, which I almost reject outright, is that humans are about as smart as it gets.

Basically, anything that causes one of the following makes anything like the singularity impossible:

* MNT isn't all that awesome
* AGI isn't all that awesome
* AGI doesn't get here before some earth-shattering technology like MNT

! An End To All Ills

So, why do I care?  Why did I write this out, besides people bugging me about it?

I'm hoping that if a Friendly AI gets to MNT before anybody else does, a genuine utopia for humanity might be possible.  What such a thing might look like is a [http://lesswrong.com/lw/xy/the_fun_theory_sequence/|big topic all by itself], but the power to re-arrange matter arbitrarily should, at least, mean the end of scarcity.

Humans actually ''need'' very little; basically just mixtures of hydrogen, oxygen, carbon, nitrogen, sodium, potassium, magnesium, calcium, sulfur, a few others.  All things that are all over the place if you can just move them around easily.

Similarly, repairing age-related damage probably wouldn't be that hard if we could actually poke at individual cell components easily.

With something in charge that is a true genius at resource management, hard moral decisions, and so on, it should be possible to really solve most actual ''problems'' that most people have.  People are still going to have difficulties, like not being popular, but ...

The world as it is now is ''horrifying''.  There are 7 year olds living off of picking food out of trash heaps.  With decent MNT and a something with a decent sense of morality overseeing it, we can get to the point where ''that'', at least, is totally impossible.  We can end death.  We can end true (involuntary) suffering and stupidity.

I hope.

! Timeline

It is natural to care a great deal about ''when'' all this will happen.  Unfortunately, the most I'm willing to commit to is "within 5 years of the point where a high-end home computer has computational power close to that of the human brain".  The computational power of the brain is believed to be somewhere between 10{SUP()}13{SUP} and 10{SUP()}16{SUP} operations per second, a quick google search suggests; there's an [http://www.transhumanist.com/volume1/moravec.htm|article by Hans Moravec] that says 100 million MIPS (10{SUP()}14{SUP}), [http://www.merkle.com/brainLimits.html|an article by Merkle] which is where I got the range I gave from, and a few others I found.  I found someone else estimating 10{SUP()}23{SUP}, which seems so far out as to be ignorable, but I'll mention it anyways.

As of this writing (mid 2010), [http://en.wikipedia.org/wiki/Instructions_per_second#Timeline_of_instructions_per_second|high end desktop processors] are doing about 100, 000 MIPS, which is 10{SUP()}11{SUP} IPS.  For the 10{SUP()}16{SUP} number, this means we need about 17 more doublings, which assuming Moore's law as always, means about 26 years.  For the 10{SUP()}13{SUP} number, it's 11 years.  For the 10{SUP()}23{SUP} number, it's 60.

In other words: wake me up when your desktop does 10 million MIPS.  That's when it'll start to get interesting.
Latest Changes

...more
Log In

Show PHP error messages
History: My Views On The Future

Source of version: 13 (current)

Latest Changes

Menu

Log In
