Interview-Vernor Vinge
From FutureNovo - Anticipating things to come
August 10, 2008
Richard Yonck, FutureNovo
FN: Dr. Vinge, thank you for taking the time to speak with me. I've read your stories and ideas over the years and enjoy them very much.
In your 1993 paper on "The Coming Technological Singularity", you talk about several possible paths that could lead to a superhuman intelligence, some entirely machine-based while others involve electronic or biological enhancement of human intelligence. Do you currently see one of these as most likely and fifteen years on from that paper, how have your views about this changed?
Vinge: Occasionally, I look at that paper to make sure of what I said there. I've been impressed by the fact there hasn't been very much change in my opinions. Of course, that can be a bad thing, but that's the way it is in this case. As to which of the four paths I listed is most likely, I think as the years go by one or another will seem more plausible. The Singularity will probably comes as some combination of different approaches, and it's possible that the transformation will play out differently than the major life changes of the last few hundred million years. Very likely this will not be a winner take all situation — you know, where after the first solution occurs, none of the others can happen. Intelligence may value thought done in different ways, on different substrates. All of the possible paths to superintelligence may be exploited. It is true that the notion of networked, embedded microprocessors seems more plausible to me now than it did in 1993. This notion of a Digital Gaia — the internet of embedded micros — is a trend that's going like gangbusters, and the idea suggests all sorts of symptoms to track. In Digital Gaia, each participating physical object is smart in the sense that it knows what it is, where it is and it can talk to nearby neighbors, in potential to the world. If there are enough such objects, it's as though the world itself wakes; reality becomes its own database. (See Karl Schroeder's notion of "thalience" in his novel Ventus.) This is quite a contrast to the centralized collection of data in server farms — which is also going like gangbusters.
FN: It seems from what you're saying and things I've thought about that the fine-grained level of sensor distribution almost emulates or is similar in certain respects to the concept of cells within a body where they have a certain degree of capability or intelligence, but certainly become far more when combined into a larger organism.
Vinge: It's a very life-like thing, almost a non-carbon based parallel to the ecosystem that already exists. The ensemble could have vastly superhuman intelligence, but with a very different flavor than our usual transhuman view of the post-Singularity world. When I need an example of how different a superhuman intelligence could be, two extreme conjecture points suggest themselves: Intelligence Amplification of our own human intelligence ("people-plus") versus this Digital Gaia notion. These are visions as different as plants and animals in the biological world.
FN: And any number of them could come about, not necessarily concurrently, but one after another or in parallel or in conjunction with one another. There could be an entire spectrum of different intelligences that are very, very different from each other.
Vinge: Yes.
FN: You've stated in your paper and elsewhere that you think the Singularity will likely occur sometime before 2030. While hardware advances have been on a pretty consistent, even accelerating track in terms of Moore's Law, do you think software advances are keeping sufficient pace to meet this time frame?
Vinge: Ah, that's the critical question for Singularity claims. I think that the Singularity is the most likely non-catastrophic event for this century and I'd be surprised if it doesn't happen by 2030. But if the software can't exploit hardware advances — put another way, if we never figure out how to put the parts together — then the hardware power may be an empty thing. So far, this question is less subject to objective argument than hardware claims. So tracking just what we are doing with the hardware is very interesting. There are various sorts of research that might give a person some added confidence — or not — about how things are going to come out.
FN: So you see it as not necessarily being as definitive or well defined a trend as the hardware advances?
Vinge: Correct. Have you read the essay that I had in IEEE Spectrum last month?
FN: Yes, in fact that's part of my next question.
Vinge: When thinking about the possibility of the Singularity, there are indicators to track. That was actually the title of the Spectrum essay.
FN: Right. "Signs of the Singularity". In it you responded to several articles published in that issue's special report on the Singularity. Does the fact that an increasing number of scientists and academics are discussing this give you hope that we may be better able to direct and create safeguards for the development of a superintelligence?
Vinge: Yes.
FN: That was very succinct.
Vinge: The question is of course, how much added confidence does it give?
FN: And how much can we really control something like this?
Vinge: There is increasingly sophisticated thought being put into safety and the meta-question of what safety means. As time goes on — until either the Singularity happens or some sort of disaster strikes — I would expect these issues to become more and more widely discussed. As I said in the Spectrum article, this may be annoying to people who are convinced that the Singularity is totally bogus or that talking about it is distracting from the things we should really be discussing.
FN: I suppose there are always people who have their own agenda and don't want to be dealing with anything else but it's certainly prudent to be thinking about it very seriously if there's even a possibility of it occurring. In terms of the Singularity, you've spoken about a soft take-off being preferable to a hard take-off which I presume most of us would agree with. But while there would be some lead up to the Singularity, doesn't the exponential quality of the concept really make a hard take-off almost inevitable?
Vinge: Maybe not inevitable, but unnervingly plausible: In real processes, exponential growth very often has complications that either terminate the exponential growth (with saturations or catastrophic collapse) or cause the exponent to change. For the Singularity, one point to make is that we have certain underlying hardware trends that are exponential. The way that manifests can be in things that are of very surprisingly different speed. So for instance you can imagine having an exponential where the things in the environment that the exponential trend are changing reach a critical point where some really extreme change can happen that has a time constant that is much, much shorter. Here the exponential trends are hardware improvement and the interesting collateral issue is the possibility of human and superhuman level machine intelligence. When/if that possibility manifests, the consequences could be much more spectacular changes and much faster change rates.
Trying to argue that the transition could be slower, that it's something that could take decades — the soft take-off version — I think that before I read Greg Bear's Blood Music, I was not into hard take-offs. Ray Kurzweil and Hans Moravec have written about transitions that take three or four or five decades. But I think in the old days I wasn't into that either. I thought that what would happen is we'd reach a certain point and there would be a runaway.
FN: I guess that's what I'm thinking with that question.
Vinge: Actually, a runaway that would happen in a year or two, that's not a hard take-off.
FN: You wouldn't see that as a hard take-off? I know you've spoken in terms of a few hundreds of hours at times in the past as a hard take-off.
Vinge: Aha yes, we may be talking about the same thing. My characterization of a hard take-off should be that whether it takes a few seconds or a year, it proceeds far faster than contemporary feedback mechanisms can cope with. Such speeds seem plausible.
As to how to justify sliding into superhuman intelligence across decades; that wasn't my original intuition but it's worth serious imagining. Here's one possible mechanism: Skeptics often say to Singularity enthusiasts, "Life and intelligence and consciousness are immensely more complicated than you imagine." What if that's correct? At the same time what if the enthusiasts about exponential improvements are also right? In other words, all the oomph of those exponential improvements is needed to meet the arguments of the naysayers.
FN: And that's what's acting as a brake or a governor or limiting factor on the process.
Vinge: Right. It's a credible scenario that it would take decades to go from the beginnings of parity to unquestionable total intellectual superiority. Thinking with scenarios is good. Scenarios should even — maybe especially — be considered for things you regard as very unlikely. For instance, pretend it's 2050 in a world where the Singularity doesn't happen and you've been asked to write an essay explaining why in retrospect that outcome should have been obvious! This exercise of the imagination gives added insight about things to watch for.
A soft takeoff, a decades-long transition with enough time for our institutions to adapt and respond — that would be very cool. Our feedback mechanisms could grow right along with the scope of the problems we face. I mentioned the writings of Ray Kurzweil and Hans Moravec. Also, J. Storrs Hall had a book called Beyond AI which is really about how to handle a long transition. In his vision of it, the early human-equivalents and superhuman-equivalents are very, very expensive sites. Very large sites. And they do not undergo an exponential runaway. They are so expensive that they are only targeted on really major issues. (As an aside, I think one thing that could slow progress for the earliest human parity hardware will be that the devices are so expensive — and so disbelieved — that they won't even be exploited as AIs. We may have to wait till Moore's Law gets these devices cheap enough for university labs and small start-ups.)
FN: So there really are a lot of factors that could influence that.
Vinge: Right.
