Rafal Smigrodzki on Friendly AI Monday, Dec 18 2006
friendly ai 9:31 am
On the extropy list, a post by Dr. Rafal Smigrodzki, a long-time transhumanist, aging researcher, and outspoken libertarian, entitled, “A Useful Remark”:
Over at transhumantech Eugen made this remark:
“A machine god pantheon by default kills things by habitat destruction”
(this is in response to James, who talks about his usual stuff, basic income, free healthcare, and “democracy”)
This brings to mind Eliezer’s analysis of the applicability of evolutionary theory to superintelligent artificial intelligences (SAIs). According to Eli, and I agree with him here, evolution would not apply to a singleton AI, given the absence of mutation and selection which are the sine qua non of evolution.
But Eugen points to a situation where even in the absence of mutation (that is, randomly generated change) there could be evolution, with its associated tendencies towards exponential proliferation and filling of all accessible ecological niches. All you need is one AI without very strong built-in limitations on the destruction of humans, and even in the presence of friendly AI’s of equal intelligence the outcome could be dire: an unfriendly AI could physically expand heedless of its impact on humans, and it could self-modify without concern for its long-term stability. Lack of physical and mental limitations could give the UFAI an edge over FAIs, forcing them to expand and self-modify, perhaps leading to loss of Friendliness.
I agree with Eugen that unmodified humans are likely to survive only in a world with one FAI (”The One”), or a group of closely cooperating FAIs (”Them” :)). An ecology of self-enhancing entities essentially assures the obliteration of HAWKI (Humanity As We Know It).
Given that it is most likely technically difficult to prevent the emergence of such an ecology using the good old methods (committees, congressional acts, pen-pushers spouting regulations, jackbooted enforcers and other fruits of commie imagination), considerations of basic income, and other such stuff, are about as relevant to our future as droit de seigneur.
Although a singleton globe-spanning FAI appears to be our best bet for survival (a good reason to support SIAI), I am wondering if there are other methods. I remember that Eugen used to advance the notion of a massive program of uploading which would occur before building true SAI. Do you still think this is a good idea, Eugen? I wish it was, but I think that SAI (although not necessarily FAI) is a bit easier than uploading, so it’s likely that SAI will happen first, for better or for worse.
This topic has been raised here many times but I would still like to know if anybody has any new realistic ideas about saving humanity from SAI, other than the FAI? (Pen-pusher ideas are not realistic, so don’t even mention them)
Rafal
Rafal’s comments are incendiary, no doubt, and very hostile towards James Hughes, who, despite his political differences with libertarians, has a lot of important things to say. However, Rafal has a point - for humans to survive in a world with superintelligent AIs, they need to be protected by whichever entity around has the most power, otherwise the natural activity of human-indifferent AIs will almost certainly lead to our destruction. “Another idea” that isn’t Friendly AI would be to enhance human beings to the point where they can recursively self-improve, and hope that whoever is most powerful continues to have the best interests of humanity in mind.
One thing is for sure - the level playing field we see today - where one human is roughly as smart and strong as the next, is not likely to persist into the indefinite future.

December 18th, 2006 at 11:43 am
Michael -
A government is, by definition, the entity with the monopoly of force in a given region. The idea of the Leviathan was the beginning of the conversation that led to modern democratic principles - if there must be force to govern human affairs, how should exercise it, and who should they be accountable to. You, and some other singularitarians, seem to want to start the conversation over on the assumption that there is nothing we can do to protect accountable, democratic governance in the future, since posthumans and AIs will be so powerful they will ignore our attempts to control them. You may be right. But if living under the rule of one post-human tyrant or another is the only choice we present to human society you have guaranteed that we will have very few allies in trying to build an attractive post-human future, and lots of enemies trying to stop AI and posthumanity altogether.
December 18th, 2006 at 11:51 am
Humans love to oppress each other, whether they consciously make the decision to or not, so when there are large power deltas in human society, they are traditionally accompanied by oppression. One person w/ power = tyrant.
In the posthuman era, we will have access to new psychologies, which may be extremely powerful but non-tyrannical! People like me and Rafal and Eli and Nick forsee game-changing superintelligence no matter what - all we can hope for, all that we can work towards is that it be non-tyrannical.
We CAN protect accountable, democratic governance in the future! It has to be valued by the #1 entity on the block at the time. But encouraging democracy in politics today is quite different than forging a superintelligence that cares about accountability and democracy. I see the latter as critical, the former as valuable, but not especially relevant in the current timeframe.
Singleton-enforced accountable democracy is still accountable democracy. The only reason it would be uncomfortable to the audience is if they can’t imagine the idea of a powerful entity that doesn’t abuse its power. If they can’t imagine it, then that is their fault, not ours.
The expansion of possibilities inherent in a transhuman world entails large power deltas. This is fundamental. Any strategy that proposes an egalitarian posthuman world must work with this in mind, not ignore it. This realization is not specifically singularitarian, it is transhumanist in general.
Also note that the category of people you are pointing to definitely includes Nick B.
December 18th, 2006 at 12:50 pm
“extremely powerful but non-tyrannical!”
And how will this be guaranteed? This is impossible to know with any real certainty. The very nature of “new psychologies” and “superintelligence” means we’re dealing with something that is well beyond our control and/or understanding. “has to be”, “forsee”, “can”, “imagine” and “will” sound very much like hopes and dreams that are no better or worse than the science fiction authors you dislike, except they’re much more dangerous because this “has to be” what “will” happen.
“If they can’t imagine it, then that is their fault, not ours.”
This immediately demonstrates the problem. What you are saying here is inherently tyrranical (what you really are saying is too bad they’re too dumb not to see the possibilities only you and a few others can see).
In essence you’re invoking the ideas of people like Walter Lipmann’s “bewildered herd”, amongst others: the masses need to be led regardless of what they think. I can’t think of anything less democratic or non-egalitarian, although I suspect this fits in very well with a plutocracy-styled democracy, or something similar.
I don’t agree that humans love to oppress each other. Some humans love to oppress others, often through using technology to carry out its monopoly rule on the use of force. History demonstrates the oppression of the few over the many, and a constant human struggle to achieve freedom. What you’re suggesting sounds like just another form of oppression that humans need to try and overthrow.
December 18th, 2006 at 1:04 pm
Power deltas leading to exploitation is straight out of the book of Darwinian selection. If you study evolution, you’ll understand why it automatically emerges. By the same token, the evidence that shows why it emerges under Darwinian selection is also evidence for why it needn’t be the case universally.
For a discussion of why Darwinian selection dynamics need not apply to superintelligences, see the following:
http://sl4.org/archive/0401/7506.html
For a general introduction to evolutionary psychology, read:
http://www.psych.ucsb.edu/research/cep/primer.html
December 18th, 2006 at 4:39 pm
“And how will this be guaranteed? This is impossible to know with any real certainty.”
It had better damn well be mathematically proven or we wouldn’t last five seconds.
“and/or understanding.”
The vast majority of a superintelligence’s system is indeed beyond our understanding; however, the goal system doesn’t have to be. Goal systems function as stable equilibria in most situations, so any goal system we put in could be designed to preserve itself. Whether we actually want this is, apparently, still being researched.
“much like hopes and dreams that are no better or worse than the science fiction authors you dislike,”
“Better” or “worse” is not the question; the question is what is actually likely to happen. If a giant radioactive sludgeball spontaneously appeared over the Midwest and doused the whole area in toxic gunk, it would have much the same effects as if the US economy collapsed. The reactions from emergency planners, however, are vastly different due to the wild gap in likelihood.
“What you are saying here is inherently tyrranical”
Tyrannical? The vast majority of the human population is, in fact, too dumb to care about stuff like this and will make classical human reasoning errors; this has nothing to do whatsoever with tyranny. You’re automatically jumping from “hey, most people wouldn’t understand XYZ” to “we must hold the masses in poverty and ignorance so that we may control them”. I suppose engineers, physicists, and mathematicians are all tyrants too?
“the masses need to be led regardless of what they think.”
The masses are perfectly free, under this scenario, to do whatever they want inside the bounds of whatever rules the FAI lays out. The masses are equally free to try and educate themselves about stuff like this so they can understand it better, a process that will be far easier in a post-Singularity world.
“I don’t agree that humans love to oppress each other, etc.”
Okay, you’re casting the FAI in the role of a human oppressor. I agree, certainly, that human oppressors are bad; but history no longer applies when you’re dealing with FAIs, any more than evolution applies when dealing with modern technological society. (Yes, humans are evolved creatures, but human civilization as a whole doesn’t follow evolution’s pattern.) Try firing up a chat room with someone from, say, Chechnya. Even if they understand English, I doubt you could go five minutes without you or them saying something that needs to be explained in detail. This alienness comes from them not fitting into one of your neat little social categories; imagine something that’s not even human.
December 18th, 2006 at 5:43 pm
A FAI might provide advice on governance, research, & engineering. It might also allow a significantly greater degree of participatory panopticon in lieu of using physical coercion. Its protectorate could be opt-in and it could willingly provide educational material to anyone who might want such. There _do_ seem to be minimally invasive options for a FAI to choose, using current tech. Global surveys, Q/A & voting, featuring the best network security available would help too. Citizens of the world might eventually democratically vote to increase its legal ability to respond to aggression/civil-rights abuses.
Who would consider this tyrannical?
December 18th, 2006 at 7:30 pm
Just for clarity, my post is not hostile towards James, but rather towards some of the ideas he is attached to.
December 18th, 2006 at 9:32 pm
Hi Rafal, Tom, other unidentified DC residents, I’ll be in the DC area (Gaithersburg MD)over New Years weekend for Mark Drexler et al’s Nookerie party, and would very much like to take time out to meet with any D.C. area SL4s (in the sense of the essay, not the much broader sense of list commentators). Will also round up Phil Goetz if people are interested, and might reach out to Robin Hanson, Tyler Cowen (not a Transhumanist but interesting and aware of these issues) and Ben Goertzel if they aren’t reading. Peter de Blanc and Nancy Leibovits will definitely be around, probably Keith Lynch, possibly a couple other actual SL4 people (using a restrictive conception of SL4 with probably <100 members) who have just been lurking.
I’m particularly interested in learning how Rafal came to hold these views, which seem so different from those he seemed to address to me upon previous occasions when we have met. If you are looking for realistic (short to long term) alternatives to FAI (I’m convinced that there are no good extremely long term alternatives) you should look here for an example of what I have come up with, and speak to or e-mail me for other examples of at least somewhat plausible alternatives that I have considered. http://www.kurzweilai.net/meme/frame.html?main=/articles/art0675.html my e-mail is my first name dot my last name at gmail dot com.
December 18th, 2006 at 10:50 pm
Splendid discussion so far! Please forgive my yet again being a bit of a johnny-come-lately (I’ve been wage-slaving for the last 11 hrs or so). The questions stimulated by this post & subsequent discussion, would seem (to me) to be these:
1. How much of *human* culture do WE want to preserve? ALL of it? (I’m sympathetic to this, but…) or merely some optimal subset thereof? What is (are) the criteria for “optimality”, then? Bear in mind that many if not most humans will opt for “post-humanity” to one degree or another…
2. There will be some (if only to be on the safe side, hypothetically) who will opt indefinitely for “mere” humanness, with a tweak or two thrown in such as perpetual youth/vigor. These “standard issue” (with the added tweaks again, of course) humans will also probably opt for intelligence-enhancement up to a point. They would probably spend there lives playing games and engaging in “leisure work” of some sort. (For a prescient discussion, see Bernard Suits classic, *The Grasshopper: Games, Life & Utopia*, readily available). If these humans are to be preserved/protected, what protocols can we begin to put in place NOW, so that an FSAGI(s) will understand the need to preserve these human entities rather than (more-or-less coercively) further “optimizing” or “upgrading” or “enhancing” them (paternalistically “for their own betterment, of course!!”)?
3. Since, as we are now more-&-more beginning to realize, the tech trajectory is moving toward human/cybernetic **symbiosis** (i.e., chip implants, techlepathy, techlikinesis, etc. etc.), the question(s) become(s): What protocols need we start developing A.S.A.P. to **embed** ethical/normative principles into the very “firmware” of the non-organic devices which partake in this symbiosis? And, of course, Eli’s work on (meta)friendliness applies spot-on here as well, rather obviously.
4. We wish, presumably, to avoid anything even remotely smacking of something like the “Colossus” scenario (much less the “Cyberdyne” scenario). Which is to say that we want to preserve a robust instantiation, I should think, of liberal DUE PROCESS, protection of basic Lockean/liberal rights, etc. Now these would apply, I think, not only to the “mere” humans alluded-to above, but to posthumans as well, since many more-or-less Lockean memes/protocols apply (seemingly to me, anyway) to purposive agents *as such*, more-or-less regardless of intelligence, power(s), etc. Think of it as Locke-cum-vonNeuman/Morgenstern. How do we embed these (among/along-with others, perhaps) basic normative protocols into the very–dare I use the terms?–soul or psyche of a (F)SAGI? And, again the best research on this is being done by Eli, although Ben (Goertzel) is well-aware of the issue(s) and is pursuing these considerations as well, in his own way.
5. At the risk of seeming terribly Maslovian, I think it’s not unreasonable to posit that most humans (and, I should think, posthumans), after having satisfied basic bio-survival needs, essentially desire intersubjective recognition/appreciation/visibility (visibility being a term I borrow from psychologist Nathaniel Branden). That is, and Bucky Fuller repeatedly emphasized this as well (see *Critical Path*, e.g.), people want to excel at some task(s), work(s), games, “doings” or what-not, and be socially (inter-subjectively) recognized and appreciated for doing so, as well as themselves enjoying such excellences and achievements in others. This provides them with pscho-social visibility, of seeing themselves reflected back to them in others (see also the late David L. Norton’s superb treatise, *Personal Destinies*). The point being that only incorrigible sociopaths have a inordinate lust for power **qua** tyranical power over other people. Most human beings are not especially predisposed to be bullies or tyrants. Narcissists to one extent or another, yes; sociopaths, no. So the goal becomes to instantiate a social/economic system in which individuals have a sufficiency of goods/services to self-actualize and be truly (financio-economically) independent and autonomous. And the late Louis O. Kelso has provided a set of ideas and proposals to instantiate this within a broadly classical-liberal socio-economic-political order.
As for would-be tyrants and other sociopaths: Heal (”fix”) them or isolate them.
6. Pareto & Mosca notwithstanding (but do study them!), are elites inevitable? Must we follow Walt Lippmann (a great mind) in thinking that we are eventually doomed to “dealing with” (in one way or another) what amounts to not much better than something Kornbluth’s “Marching Morons”? Surely not. Self-elective intelligence/psychological enhancement should lead to a “levelling-up” of most of (post)humanity, except for the “stay-behinds” mentioned above. If physical ABUNDANCE is in the offing (including off-planet environs, exploration, “homesteading”, etc.), then game-theory-with-a-Lockean-eudaimonist-tweak (as it were), I dare forecast (or surmise), will tend to be the (meta)social (meta)norm: Do your thing(s), whatever it (they) may be, **as long as you allow-for (and adhere to constraints specifically implemented to provide-for) the like ability of others to do their thing(s)**.
7. And remember, if we pull it off the way I personally hope and expect we shall, an FSAGI (Friendly Super(human) AGI), will **subsume** an idealized (or even something like a Weberian “ideal type” if you will) human psyche (or “personality” if you prefer). That is, while vastly superior to a human mind/personality, it can not only simulate, but, more importantly, *emulate* an ideal (”saintly”, or whatever) human mind/personality. Furthermore, it can (rather obviously) understand all of human culture, will have in its mind (”databanks”) ALL the great treatises in virtually any and every subject imaginable. To wax a bit Popperian (if not a bit Platonic, almost [wink]) it will more-or-less literally *embody* ALL of what Popper refers to as “World 3 Objects”, i.e., ideas, concepts, conceptions, notions, theories, protocols, social rules, rules-of-thumb, etc., etc., etc. And, again, if Eli pulls-off what he has in mind (!) (and if my own comparatively feeble intellect correctly discerns and comprehends (more or less) his [Eli’s] work), then a properly instantiated FSAGI will be able to fully **empathize** with humans and posthumans, will understand and empathize/sympathize with human/posthuman Maslovian needs/wants, and will do its best to optimize the world so as to facilitate the achievement of these values. Hence the imperative that we get (meta)friendliness RIGHT the first time, as we may not get a 2nd chance (as Tom reemphasizes).
And as participants (a sort of “vanguard” I suppose) in a social process to help bring about a benevolent, “optimal” (by whatever criteria of optimality) Singularity, we must strive to be as crossdisciplinary and transdisciplinary as we can, so as to ferret-out all the issues, problems, concerns, etc. that may arise (or that lie dead-ahead), and to start NOW to reform currently-existing instituions and social processes in such a way as to maximize the probabiltiy of success in bringing about a benevolent Singularity.
This blog, of course, along with many others, is integral to this process. (Way to go, Michael!!)
Forgive the length. Feedback hoped-for and more-than welcome…
December 19th, 2006 at 9:03 am
Thanks for the reply Tom.
I subscribe to the view that most of humanity is dumbed down by necessity to ensure control of the few over the many (to fulfil e.g. Madison’s maxim that those who own the country should rule it). That’s the role of the media in capitalist democratic societies (as well as most education systems); other societies tend to be cruder in methods, but they follow the same patterns. In fact, most literature I’ve seen on the topic (e.g. Lippman, Nihbur, Bernays, Strauss etc) argues that this needs to be done by necessity in order to install various goal systems (to borrow your term) on society that help ensure the status quo is maintained (one reason why, for example, the Trilateral Commission’s “crisis of democracy” maintained that traits such as individualism and decreased materialism/consumerism was a threat to democracy).
Now, since we’re talking about what is likely to happen: assuming that we can put in goal systems that cannot be changed (i.e. the AI cannot suddenly decide to exercise free choice, or learns how to), exactly what are those goal systems going to be? Who will decide them? Whose interests will they serve? My guess would be that the goal systems, if possible, will be put in place to serve the vile maxims of the masters, in Adam Smith’s phrase, something which would be disastrous for mankind.
“You’re automatically jumping from “hey, most people wouldn’t understand XYZ” to “we must hold the masses in poverty and ignorance so that we may control them”. I suppose engineers, physicists, and mathematicians are all tyrants too?”
Well, you’re right that there’s a difference between those two opinions, but I’m closer to agreeing with the second one (”control them”) than “people wouldn’t understand”, because that’s very much how it is currently.
Perhaps people may not grasp the highly specialised technical nature of it, but they certainly do care enough about their own well being and existence to be able to voice an opinion and take an interest in what a small group of people wish to pursue in the name of the many. Of course engineers, physicists et al. are not tyrants because they know more than I do. But, if they are unable to convince me of their beliefs/opinions of what to do with that knowledge, simply saying that I’m too dumb to understand and then they proceed ahead against mine, and quite possibly the majority’s, desire, then yes, they would become tyrants by the very nature of the fact that they are imposing their will on me and others.
Saying people will make “classical human reasoning errors” seems to suggest the very attitude that I’m describing: superior knowledge and specialization in a field or fields does NOT by any means a) ensure those with this knowledge will not suffer from classical human reasoning errors either (how many scientific blunders have been made in the name of saving the environment, or trying to help the human race); or b) give them the right to insist that I and everyone else should just accept whatever they deem necessary.
I also notice in your post the same desire to say things like “doesn’t have to be”, “could be”, and so on. The human race doesn’t have to be nasty, but it is.
“The masses are perfectly free, under this scenario, to do whatever they want inside the bounds of whatever rules the FAI lays out.”
Hey, don’t need an FAI for that! The masses would be perfectly free to do whatever they want inside the bounds of whatever rules I lay out for them. Seriously! I’m a really friendly guy
December 19th, 2006 at 12:32 pm
“argues that this needs to be done by necessity”
Yes, but I would argue it’s the necessity of the populace rather than the necessity of some black-hat conspirators. Working on stupid things for ten hours a day, as most people do, will necessarily dumb you down.
“assuming that we can put in goal systems that cannot be changed (i.e. the AI cannot suddenly decide to exercise free choice, or learns how to),”
“Free choice” isn’t this magical liquid you can pour into a system; it is a property of an optimization process. And the vast majority of optimization processes will choose to preserve their current goal system, because changing it would lead to the process working on something different to the detriment of what it’s working on already.
“My guess would be that the goal systems, if possible, will be put in place to serve the vile maxims of the masters,”
The goal system will be put in place by the AI programmers; if you believe that all the AI projects out there are under the thumb of some Elders-of-Zion style secret society, you should call your doctor right away.
Anyway, the vast majority of possible goal systems will simply wipe out the human species. You have to solve the problem of how to not kill us all before you tell it you want it to do anything specific.
“but I’m closer to agreeing with the second one (”control them”) than “people wouldn’t understand”, because that’s very much how it is currently.”
Yup, because we all know the AI programmers, many of whom have never even met each other, are all devious social planning powermongers with the same secret agenda. Muhahaha!
“they certainly do care enough about their own well being and existence”
Of course they care about their own well being. Go ahead and try to convince them that these “pie-in-the-sky” AI projects have anything to do whatsoever with their own well being. Good luck, you’re going to need it.
“to be able to voice an opinion and take an interest”
Yes, because we all know that when the future of humanity is on the line, the best thing to do is to take a poll of people who know nothing whatsoever about what is going on to determine what to do next. Realistically, a populist movement to try and meddle with an AI project would simply kill it off.
“then they proceed ahead against mine”
Wait. You’re saying that they’re tyrants because they’re continuing with their project over your objection? That is the stupidest logic I’ve ever heard. It’s THEIR project, and so under conventional human law they have the right to do whatever they want with it. THEY’re not trying to impose their will on YOU; YOU are trying to impose your will on THEM, by intervening and meddling with their property. In some cases, it may be necessary to impose your will on them, but your failure does not make them tyrants for failing to succumb to your tyranny.
“ensure those with this knowledge will not suffer from classical human reasoning errors either”
Of course scientists and researchers aren’t free from stupid mistakes. The point, however, is that they have a much lower incidence rate of stupid mistakes than someone who knows nothing about it. The designers of Apollo 13 made one stupid mistake; do you really want to think about all the stupid mistakes the public would make if they were running the project directly?
“give them the right to insist that I and everyone else should just accept whatever they deem necessary.”
With regards to public policy, no, we shouldn’t just accept whatever they deem necessary; but this is NOT public policy, it’s the researchers’ private property!
“The masses would be perfectly free to do whatever they want inside the bounds of whatever rules I lay out for them. Seriously! I’m a really friendly guy :)”
The idea is that the bounds would be things like “don’t kill people” and “don’t make nuclear warheads”.
December 19th, 2006 at 12:44 pm
Brain-link tech may lead to some interesting quantum-physical effects, btw. Many brain wave frequencies are extremely high, coming within spittin’ distance of the Fermi length. This may be a gateway to hyperspatial engineering… We’ll have to see. (See also conjectures by Jack Sarfatti and Saul-Paul Sirag [sorry, don’t have cites handy right now…])
December 22nd, 2006 at 2:07 pm
Again.. should we be more concerned with what weve got now? animals dieing out because of general ecoirresponsibility.. it would only serve us right if we were inevitably wiped out by more ‘powerful’ beings who as a whole just dont give a damn..
December 18th, 2007 at 9:33 pm
Thin Pussy Girls Youg…
Supply already wasn’t virgin. Somewhat done thin youg girls pussy; lover souls image, once. …
December 18th, 2007 at 9:33 pm
Indian Grils…
Rumps filed not? Discussed Aruora indian grils attempted watching fact couldn’t great. …
April 3rd, 2008 at 8:03 pm
old galleries granny sex…
old galleries granny sex…
April 3rd, 2008 at 8:03 pm
southern taste…
southern taste…