Kurzweil Claims That the Singularity Will Happen by 2029
futurism.com
futurism.com
1) Choose a recent 5-year trend, preferably one increasing geometrically.
2) Extrapolate out this trend n years.
3) Your prediction is either the number from (2) above, or, for bonus points, some directly dependent result (e.g., in 2000, connection speeds were increasing, instead of predicting "connections will be X mbps", the prediction is something like "we will be able to watch HD video in real time." This sounds more impressive.)
4) Good predictions need a timeframe, so choose 2n to give yourself some buffer for being wrong.
5) You will be a highly sought after futurist at this point, so please send 1/2 your book royalties and speaking fees to my address. (PM for details)
Look at his predictions from the 1989 timeframe and count up how many were just extrapolating Moore's law and coming up with different implications.
My point was that making unsurprising, long-term predictions (10-20 years) is both easy and useless. But for some reason we call people who do this, "futurists".
Now there's another name for people who can make surprising, short-term predictions (< 10 years). We call them "rich".
Source: eaten with him, carried his luggage.
The cynic in me says that 2029 is still within bounds where Kurzweil can still participate. He would be 81 then.
The guy is a brilliant polymath but I get the impression he thinks he can cheat the big sleep.
Frankly, it perturbs me when people dump on others who have the courage to make predictions or misrepresent them. Don't get me wrong, I despise people who are consistently wrong with their predictions and have the audacity to pawn themselves off as a visionary.
Rather than dump on somebody else's predictions, I'd encourage other people to make their own and stake something on it. Put some skin in the same. Where do you think we'll be in 2029? Why?
Have you considered that some people think it's stupid to make predictions like this so they wouldn't want to make their own predictions?
It's like telling people criticizing numerologists to perform better numerology to prove something.
Making predictions and providing analysis for those predictions is actually incredibly valuable because it helps people and companies anticipate trends and hedge their strategies.
Unlike numerology, making predictions grounded in reason is actually a valuable and timeless skill that will benefit most people. I'm not encouraging people to engage in nonsense.
Except realistically, how is this supposed to happen? Just because somebody runs ./ai_overlord on some linux box, or some VM cloud, doesn't mean that this AI suddenly somehow magically has insane industrial capacity. It can't necessarily network with all the physical infrastructure and vehicles to start carrying out goals. It's existence is limited to the various RAM chips of a digital network. It will take much longer before an AGI can actually even have a major physical impact on the world.
Also, intelligence is not magic. There are certain mathematical limits to intelligence. An AGI cannot solve the halting problem or compute the Kolmogorov complexity of some data. It can't suddenly build any physical object from raw materials without a huge amount of infrastructural work being setup in advance.
And let's not forget that human beings are also Turing complete. Which means, in theory, anything an AGI can do, we can also do. The main difference is that an AGI would be able to computationally solve problems at much greater speeds than a human. But it's totally unclear to me if this somehow translates into a sudden complete overhaul of all of our civilizational infrastructure overnight. Frankly, that doesn't seem likely. I think it's more likely the first AGIs will just trade derivatives or recommend Netflix queues.
The AI singularity could happen before the inflexion point of the S-curve, or some time after, or never. And as others have already commented, it depends on how you define "singularity". So I am not holding my breath waiting for the rapture.
Side note, "certainty" is a feeling. If we were given extra senses outside of our own brain and given a sense of certainty this was ourselves, we would never know if we were controlled by the AI itself, similarly to how the brain stem hosts the higher brain functions such as the prefrontal cortex. Random, sorry.
I think reaching the singularity will require a few more optimization tricks, as we saw with pretraining many layer networks, and also a breakthrough or two in fast hardware, associative memory, etc.
I have been working in the field of AI since 1982, and I find it easy to both admit how much I don't know and that I have a lot to be modest about. With these caveats, I predict that AGIs to be hybrids of deep learning neural networks and more conventional AI systems.
He has impressive technical work to his name, but his written predictions are wishful thinking.
But in practice? What if there's an asymptotic limit to intelligence? ie. We can get really smart, but there's only so smart we can get. What if we just get faster?
You could set an average person a really hard math exam and maybe they'll get 50% correct over the course of 2hrs. What if all we get out of AI is a computer getting 50% correct in 2 seconds?
Anecdotally, this is also my general impression of the state of AI as it exists today. Our big advances are generally just making simple calculations much much faster than before. Deep blue, AlphaGo, Google Search, Deep Learning in general; it's not smarter, just faster.
More formally it's the acceleration of machine intelligence and subsequent capabilities to such an extent that Star Trek would look downright ancient compared to this post-singularity reality (save for the warp-space travel part, but interestingly enough a "decent AI" would probably deem it much more prioritary than humans to colonize outer space in order to ensure its own survival, and hopefully ours as well if our relationship is one of cooperation/parasitism).
It is a fascinating concept, but it depicts such an unprecendented discontinuity in history, a "civilizational breakthrough" that's so dramatic in scope, that there's no historical ground to it whatsoever over ten millenia. Which makes the concept of singularity somewhat of a belief, much less "not a question of if but when" than, say, self-driving cars, quantum computing or even the human-level AI threshold itself.
Nock Bostrom "Superintelligence" book is a bit tedious and descritive, but I think it does a much better job than Kurzveil's rather pop-oriented publications (though I praise him for helping tech be known and sought after by the general public, it's exactly what we need to scale).
AI is coming on a vastly different 'hardware platform'. There is little reason to believe that it will be limited by cognitive constraints implicit to the human form. Extrapolating the rate of growth in processing power makes it extremely likely to surpass human scale processing very soon.
What both these things imply is that AI has little to gain by 'tieing' itself down to humans through a neural lace or the neocortex.
Humans can provide the drives on an ongoing basis. The danger is providing any drive that can get out of hand before we can revise it.
All programs have a very explicit goal - a lot more specific than most humans do. You will have to try very hard to write a program to not have a goal.
Your laptop has a need of battery life. It lets you know of the same when it is out of it. Not very different from humans being hungry.
AI might not have hopes or dreams. But a lot of recent game playing AI can simulate possible future scenarios and develop responses for them. Not very different from hopes and dreams.
If you write a C program, in some ways, the only thing in the 'programs mind' is to finish its goal and terminate itself. In a hurry to finish our goals and terminate ourselves is also perhaps how an Alien intelligence would describe the human race looking in from the outside.
Humans can provide the drive only because we make bigger plans than AI. And our plans are a direct expression of our computing power. There is no good reason why AI wont beat that processing power.
At the same time, giving an AI a goal is the moment when we become subject to an indifferent genie who doesn't share our values.
In terms of your original question "what use would they have for us?" I guess my answer is something like: if we program them correctly, helping us will be their only goal. If we give them other goals, even slightly off the mark and don't have a tightly integrated way to course correct (like the neural lace), then we will be like ants crushed by an indifferent boot-heel.
Even if we assume that we have a way to 'program' them, the problem is that we ourselves dont know what it means to 'help' humans.
If those humans were a part of the US military facing a situation similar to ww2, it would mean that AI 'help' those humans drop some nuclear bombs. If those humans belong to a religious fundamentalist group, it might mean that the AI help further the cause of their gods.
One probable outcome, that this chain of thought leads to, is that perhaps the last AI standing will be the one which kills all other AI, or has the capability to. A possible corollary being that to do so the last AI standing will be the one which is able to prioritise its own survival over everyone else - even other humans. Which was not the plan that we started with.
On the other hand, a Rhoomba can seek out power to stay alive, probably better than a 2 year old.
The human form evolved in an environment of scarcity. Early humans had to provide for food three times a day, and often risked their lives to procure it. On the contrary, the average AWS instance almost never shuts down for lack of power.
We might not have a better model right now, but it seems extremely unlikely that AI will look anything close to the model we are talking about.
Both of those things ended up happening a couple of years before 2000. While practically the entire world was saying that Kurzweil's far too optimistic, it turned out on the whole they were slightly too pessimistic.
It would be foolish to discount this.
Kurzweil has done great work mapping out when AI is likely to reach each of these milestones and so far has been far more successful in his public predictions than anyone else.
[1] http://scienceblogs.com/pharyngula/2010/08/21/kurzweil-still...
It's a fairly recent development and I'm not sure how many futurists have refitted their models for that.
My own two cents thrown into the ring say that the class of future portrayed in the Singularity is Near (TSiN) is something of a foregone conclusion. It's quite likely that we'll all be wildly, humorously wrong about the details of implementation, culture and usage, but - barring existential catastrophe or disaster - the technological capabilities discussed in TSiN will come to pass. The human brain will be reverse engineered, simulated and improved upon. The same goes for the human body; radical life extension is one desirable outcome of this engineering process. We will merge with our machines as nanotechnology and molecular manufacturing become mature technologies. Recursively self-improving general artificial intelligence will develop, and then life will really get interesting very quickly. And so forth ... the question is not whether these things will happen, but rather when they will happen - and more importantly, are we going to be alive and in good health to see this wondrous future?
We humans are in the process of building tools that enable us to create or meaningfully interact with ever-greater complexity. There is one important area of complexity management in which we seem to be making little headway, however: the organization of humans in business and research. For all that we can now accomplish with faster computers and enormous leaps in telecommunications, we don't seem to have made significant inroads in getting large numbers of humans to cooperate efficiently. I am prepared to go out on a limb here, as I have done before, and say that business and research cycles that involve standard-issue humans are incompressible beneath a certain duration - they cannot be made to happen much faster than is possible today.
This is not to say that they cannot be made cheaper. But cheaper doesn't equate to faster business and research cycles; rather, it means that any given problem will be tackled by many more parallel attempts. Expensive projects mean conservative funding organizations, which means organizational matters proceed at a slow pace. This is a defining characteristic of our time: we have blindingly fast rates of research and technological advances once the money is on the table, but the cycles of business, fundraising and research are still chained to the old human timetable.
Kurzweil's Singularity is a Vingean slow burn across a decade, driven by recursively self-improving AI, enhanced human intelligence and the merger of the two. Interestingly, Kurzweil employs much the same arguments against a hard takeoff scenario - in which these processes of self-improvement in AI occur in a matter of hours or days - as I am employing against his proposed timescale: complexity must be managed and there are limits as to how fast this can happen. But artificial intelligence, or improved human intelligence, most likely through machine enhancement, is at the heart of the process. Intelligence can be thought of as the capacity for dealing with complexity; if we improve this capacity, then all the old limits we worked within can be pushed outwards. We don't need to search for keys to complexity if we can manage the complexity directly. Once the process of intelligence enhancement begins in earnest, then we can start to talk about compressing business cycles that existed due to the limits of present day human workers, individually and collectively.
Until we start pushing these limits, we're still stuck with the slow human organizational friction, limits on complexity management, and a limit on exponential growth. Couple this with slow progress towards both organizational efficiency and the development of general artificial intelligence, and this is why I believe that Kurzweil is optimistic by at least a decade or two.