Making Sense of the Singularity Thursday, Jul 16 2009
friendly ai and singularity 10:56 pm
Mike Treder, Executive Director of the IEET, who will be attending the upcoming Singularity Summit 2009 in New York on October 3-4 (register now!) has commented on the whole idea of the Singularity over at the IEET blog.
Like many Singularity commentaries, Mike’s begins by conflating three totally different concepts that all have unfortunately come to be called “Singularity”, depending on who you’re talking to — the idea of “a theorized future point of discontinuity” (Event Horizon), “when events will accelerate at such a pace” (Accelerating Change), “that normal unenhanced humans will be unable to predict or even understand the rapid changes occurring in the world around them” (Event Horizon again). Then, according to the IEET Encyclopedia of Terms and Ideas, Mike writes, “It is assumed, usually, that this point will be reached as a result of a coming “intelligence explosion,” most likely driven by a powerful recursively self-improving AI.” (Intelligence Explosion.) Three definitions in one — not good for understanding what you are analyzing. If the convention includes all three definitions, the convention should be discarded.
Later on, Mike writes:
Insisting on the possibility—or, even more strongly, asserting the inevitability—of an uncertain and debatable but incredibly momentous event leaves proponents vulnerable to a charge that they lack rigor and discipline in their thinking, that they have fallen prey to hopefulness and shed any semblance of healthy skepticism. If they cannot restrain themselves from heartily endorsing an unprovable proposition, than what credibility have they for other declarations or recommendations they might make?
This is a slightly odd paragraph. Firstly… “insisting on the possibility…” of an event? A possibility can be arbitrarily low probability — it’s not very insistent. If there is no possibility, then there is no need for debate. Remember, it isn’t even clear which possibility is even being discussed — is it the Event Horizon, Accelerating Change, or the Intelligence Explosion?
Indeed, no Singularitarians should assert the inevitability of the Singularity. 10 years ago, 5 years ago, and today, only a small minority even do. Ray Kurzweil seems to be one of the very few Singularitarians that even implies it. To say just a couple small things in order to dispel other possible myths about Singularitarians:
1. The Singularity is not necessarily a good thing, a “Singularity” could mean that everyone dies at the hands of a recursively self-improving AI or upload that is indifferent to our welfare.
2. The Singularity is not at all inevitable. We could easily blow ourselves up first.
Other Singularitarians that agree with the above two (seemingly obvious) statements, feel free to say so in the comments if you consider it useful.
Mike writes:
More troubling is the suggestion by some (all?) singularitarians that the outcome they seek is not only possible but desirable.
To me, the Singularity is nothing more than creating greater-than-human intelligence… oddly enough, this is Vinge’s original definition. Is greater-than-human intelligence necessarily desirable? Absolutely not. If such intelligence(s) have values indifferent to humanity, we could all die, as we’ve emphasized again, and again, and again. How many more times do we need to say it?
Mike writes:
Given the substantial amount of uncertainty—which they themselves admit—surrounding the nature and impacts of such an occurrence, it seems imprudent to stamp the Singularity as unquestionably “a good thing.”
In “a good thing”, he links The Singularitarian Principles as a source. But the top of the page says, “This document has been marked as wrong, obsolete, deprecated by an improved version, or just plain old.” Really, truly. Yudkowsky, the author of that document, dropped the idea of the Singularity as an unqualified “good thing” around 2002. He’s mentioned this in practically every single talk since then. Every “Singularitarian” mentions this practically every time they talk about the Singularity. Are we Pollyannish optimists or psychotic doom-mongers? We can’t be both simultaneously.
Mike writes:
Worse yet, some who proudly say they’re working to bring about the Singularity have the temerity to proclaim that they alone hold the keys to making it a “friendly” event.
Absolutely not! If anyone ever claimed that, they were wrong. We (as in the SIAI) have only claimed that we focus on the issue of making AI friendly the most enthusiastically of any group working towards AGI, which is true. But we are in contact with many other groups that care about friendliness too, and that is wonderful. It’s a research field, which was pioneered by our group, but we can only hope that it will continue to extend far beyond its boundaries, as it has been. (Much more successfully than Mike’s belief that attempts to engineer friendliness in seed AI are hubristic, thankfully.) People at Google care about it, people at Hanson Robotics care about it, many independent researchers care about it, and many institutional academic researchers care about it. So, far from claiming that we alone hold any exclusive keys, we welcome anyone to contribute to the field. Please, if you have the funds and the intelligence, quit your day job and work on Friendly AI — that would be great.
Mike writes:
Besides sounding childishly naive, such a claim also invokes the specter of technocracy: if only all the big issues of this world were left to the few really smart people to solve, everything would turn out fine.
The need to work towards Friendly AI comes from simple assumptions:
1) Some possible future superintelligences could kill us all, as almost happens in The Matrix and Terminator.
2) Some possible future superintelligences could make the world a lot better, as is portrayed in some sci-fi books.
3) The first seed AI to cross a given threshold could bootstrap itself to superintelligence.
4) The way we design that seed AI could influence the way it subsequently develops.
The same points are made, of course, in an essay by Nick Bostrom (coincidentally, the Chair of the IEET, Mike’s host organization), “Ethical Issues in Advanced Artificial Intelligence”:
The option to defer many decisions to the superintelligence does not mean that we can afford to be complacent in how we construct the superintelligence. On the contrary, the setting up of initial conditions, and in particular the selection of a top-level goal for the superintelligence, is of the utmost importance. Our entire future may hinge on how we solve these problems.
Who will construct the superintelligence “seed”? Probably a relatively small group, perhaps a few thousand people at most, more likely a few dozen. Is Nick Bostrom “childishly naive” because he acknowledges the importance of initial motivations in a superintelligence and says that the “entire future may hinge on how we solve these problems”? This is one of the only times I have ever seen the Executive Director of an organization criticize the public statements of its Chair so explicitly.
Mike writes:
It’s implied, moreover, that meddling democrats and pesky government regulation will only slow things down and might even prevent the smart singularitarians from saving the day.
Negative. Where is this implied? It is extremely unlikely that any significant portion of society will take the challenge of superintelligence seriously before it is too late, anyway. This is not a political statement, it’s simply a logistic forecast.
It’s not an all-or-nothing thing. Participants in democratic politics (like myself) can analyze and vote on issues, government regulation can be responsibly formulated, and singularitarians can work on AI friendliness. Why does there have to be any inherent antagonism between these motivations? I’m a living example that someone can care about all of them at once. You can too.
Mike writes:
Those who promote the idea that a Technological Singularity is not only possible and desirable but that its advent can be hastened through our efforts must be aware of the obvious parallels between their own beliefs and those of Christian Millenarians.
Does the same apply to those who think that nuclear fusion is possible and desirable and its advent can be hastened through our efforts? Why not — nuclear fusion promises a huge abundance of cheap power, and its proponents think it could render fossil fuels obsolete — a “superlative” view if there ever was one. Actually, it seems like many, many discussed technologies, if not the majority, are considered possible, desirable, and hasten-able through the efforts of the researchers working towards them — if not, then why would anyone bother with them?
Where is the parallel with Millenarians? There are so many differences. You can read this page for a summary — it’s the first Google result for “Rapture of the Nerds”. I am awaiting a rebuttal to the points in that post — if Singularitarianism is really parallel to Millenialism, then shouldn’t it be easy to knock down those arguments?
Mike writes:
Finally, the proposal that a Singularity can be managed for “friendliness” seems hopelessly hubristic.
Again, your chairman, Nick Bostrom, argues this. Why be the Executive Director of an organization whose Chair is “hopelessly hubristic” or “childishly naive”?
Bostrom writes:
It seems that the best way to ensure that a superintelligence will have a beneficial impact on the world is to endow it with philanthropic values. Its top goal should be friendliness. How exactly friendliness should be understood and how it should be implemented, and how the amity should be apportioned between different people and nonhuman creatures is a matter that merits further consideration.
Friendliness in AI is necessary — the alternative is building it according to some non-friendly optimization criteria. And who could argue that an altruistic seed AI or nascent superintelligence wouldn’t be more likely to develop into a benevolent superintelligence than one built without the “hopelessly hubristic” attitude that Mike denigrates? The attitude that morality in AI is necessary is not unique to Singularitarians — it is shared by serious academics like Wendell Wallach and Colin Allen, the authors of Moral Machines.
Would you rather have a seed AI based on Gandhi or Hitler? Is it “hopelessly hubristic” to think that a seed AI (or enhanced human intelligence) based on Gandhi would be much more likely to lead to a positive future for humanity than one based on Hitler? Morality is inherent to the structure of the brain — it must exist as motivations in the cognitive architecture of the agent. It will not be instilled through mere osmosis, as it is with some children. A psychopath can spend a thousand years around the nicest guy in the world and not change a bit, because his neural structure is that way, and barring neuroengineering, that isn’t going to change.
Perhaps, the attitude that friendliness is necessary is actually required to ensure our future survival, as Nick Bostrom concludes in his paper:
It seems that the best way to ensure that a superintelligence will have a beneficial impact on the world is to endow it with philanthropic values. Its top goal should be friendliness.
Pretty simple.




Thanks, Michael, for your response to my article. My hope is that we can discuss these matters civilly and rationally, and that all of us will come out with a deeper understanding of how we can work toward beneficial outcomes.
Let me make a few quick points:
Just because Nick and I are connected with the same organization does not mean we agree on every point of view; we’re awfully democratic in that way.
If you can’t see the obvious and clear similarities between Singularitarianism and Millenialism, you’ll have a really difficult time trying to explain the differences.
The technocratic leanings and hubris I object to are based on what I’ve observed over nearly ten years of interaction with people in the SIAI fold. Granted, some of those attitudes seem to have moderated, but I submit there is still a long way to go.
My aim in writing the article you’ve criticized is to shine a light on what I see as the shortcomings of dogmatic singularitarianism; if you and your colleagues can demonstrate that I’m totally wrong, I’ll be quite pleased. That would mean a lot of progress has been made.
Finally, do you disagree with the recommendations in my final paragraph? Because that’s the main point: we may have different views on the history of singularitarian thinking, but I’d like to identify and emphasize where we agree and can support each other’s worthy efforts.
Hi Mike, thanks for the comment. I hope I’m being relatively civil, if fussy.
To respond to your points:
It would make sense for you and Nick to disagree on some things, but we both know that Nick is very intelligent and careful in his thinking and it seems highly unlikely that he would make the childish or naive errors you imply. There must be other reasons he supports friendliness besides hubris.
Of course I can see similarities between Singularitarianism and Millenialism. It’s just that that line of thinking seems to lead too often to superficial dismissals of Singularitarianism without understanding the reasoning behind a lot of it. For instance, attributing a desire for AI friendliness to hubris or libertarianism instead of deriving from work in moral philosophy (Greene, Hume) and cognitive science (Pinker, Cosmides & Tooby, etc.)
I understand that you have quite a bit of experience interacting with people in the SIAI fold, which makes it all the more unusual when you imply that Singularitarians think (in general) that any Singularity would be a good thing, even if an obsolete document from 2000 says so. (That part you quote is probably the most glaringly obsolete part in the entire document, too.)
I understand your aim in writing the article… naturally, my aim in writing this blog post is to criticize some of your criticisms.
I actually disagree with a lot of the tone in the final paragraph. What to me looks like, “the natural benefits of benevolent AI superintelligence if we do it right” probably looks to you like “unbridled techno-optimism”. My great concerns about the lack of friendliness awareness among AGI researchers would probably appear to you as “cynical techno-pessimism”. I consider the possibility of a Singularity well-substantiated, and continue to display “healthy skepticism” towards it, but I’ve applied my extensive skepticism to the concept and it has survived.
It’s worth pointing out that I have no personal reasons to wish that a Singularity would happen. A long, slow “Surge” would be much more preferable to me. I am in no huge rush, and my life is quite great as it is right at this very moment. I am the opposite of the nerd rapturist straw-man painted by folks such as Charles Stross. Yet I have been accused of being a techno Rapturist for about a decade, so naturally I will dispute the arguments I’ve heard hundreds of times that are most often associated with such accusations.
You write about “resist(ing) the siren call of technocracy”. Well, again, we have disagreement. I submit to the “call” of superintelligence, just as other innovators submit to the “call” of developing renewable power sources or whatever technology you care to pick. Superintelligence is unique in that, if successfully developed, it would not be just a technology, but actually an intelligent agent, so wanting superintelligence to help us make hard decisions is not technocracy at all, but similar to hoping that future humans will help us make better decisions.
You say, “if they will recognize that nothing is certain”. It is a strawman to say that Singularitarians think the Singularity is certain, except for maybe Kurzweil. You know this, so why pretend otherwise? The Singularity doesn’t have to be certain to be an absolutely huge deal, just like nuclear war doesn’t have to be certain to be a huge deal.
You say, “working to relieve suffering today is every bit as important as chasing the promises of future technological potential”. My response to this is to refer you to the Transhumanist FAQ, specifically the question “Shouldn’t we concentrate on current problems …” If we solve poverty and aging and then all die at the hands of a human-indifferent superintelligence, then all that prior work would be pretty useless, wouldn’t it? Conversely, working on friendly superintelligence can accelerate progress towards the goals of eliminating poverty and aging more effectively than any other strategy, in my opinion.
This argument between us (originally beginning between me and James Hughes on wta-talk quite a few years ago) is nothing new. If it were a newborn when it started, it would probably be in kindergarten by now. Our contradicting beliefs and emphases inevitably bring us into conflict. We compete over the same slice of memespace and the minds of much of a common pool of people, and the conditions for such conflicts to occur will naturally be created again and again. There is no way out of it. ;) We can be relatively civil, but obviously I am going to be slightly annoyed when my belief set is called “childishly naive”, despite years of careful thought and self-questioning on my part.
Memetic chariots, to war! ;)
Just FYI: Someone at SIAI should perhaps do a bit of editing on this page –
http://en.wikipedia.org/wiki/Singularitarianism
You and I might stipulate that Eliezer’s 2000 thinking is now obsolete, but the “wisdom of the crowds” apparently hasn’t got the message yet.
I generally encourage thinking people not to waste their time editing free encyclopaedias. Who cares what Wikipedia thinks? That site seems to be maintained by teenagers as homework. I’m sure the SI has better things to be getting on with.
Mike, the problem with that definition is that if you replace “Singularity”, with “good Singularity”, it all makes sense — of course we think a good Singularity is possible by definition. There has been definitional drift whereby “Singularity” has come to mean “good Singularity” by default, with a “bad Singularity” being where everyone dies or there is some tyrannical singleton. Sure, the Wikipedia page can be misleading, but I thought you’ve all of these arguments and issues many times.
Michael,
I self-identify as a Singularitarian and both appreciate your defense of the idea as well as agree in principle with all of your points. Having said that, I would like to offer a mild critique on the importance you place in distinguishing between the three major definitions of the Singularity. You have written very intelligently about this before and, though I accept your distinctions between the definitions, I can’t help thinking your emphasis on these distinctions is somewhat misplaced. Molecular manufacturing, radical life extension, synthetic biology, etc. can all represent discontinuities in our ability to predict or imagine the future that are potentially just as profound as smarter than human intelligence. It seems to me that, for better or worse, the Singularity meme has come to include any future technology that has the potential for major changes in human culture and I think that, by and large, that’s an accurate reflection of the general public’s perception. Granted, each of these ideas differ in their likelihood, risk to humanity, timeline for development, and corollary issues and deserve to be discussed individually but the same could be said about the differences between uploading and building AI from scratch. It all seems very similar to the debate as to whether or not alternate history should be considered a sub-genre of sci-fi or not, valid but ultimately niche and irrelevant to the majority of people likely to encounter the meme for the first time. Also, I would hate to see Aubrey De Gray who, though I could be wrong, seems unlikely to speak on superhuman intelligence excluded from future Singularity Summits ;).
Probably the wrong post to comment on this, but you seem to have such a problem with Singularity having different definitions, which leads to at least some of this debate.
You point out how Treder jumps between 3 definitions:
-Event Horizon
-Accelerating Change
-Intelligence Explosion
I think these are pretty complementary, rather than confusing. It is a difficult concept to wrap ones head around, by it’s very nature, but the concept that is really important is a mix of these three.
Accelerating Change brings about the Intelligence Explosion, and as that happens our comprehension of what comes next breaks down. We hit an Event Horizon.
Your work towards the Intelligence Explosion is aided by the Accelerating Change, and is driven by our natural desire to comprehend what is beyond the Horizon.
To understand all three is to understand what we’re all aspiring for.