On the Impossibility of Supersized Machines
arxiv.org
arxiv.org
It opens by parodying the classic futurist parable of history as exponential growth in intellect and technology, recasting history as a tale of growth in size (after all, the earliest organisms were tiny). It recasts fears about superintelligent AI as fears about super-large beings (wouldn't you be afraid of something that could step on you as easily as you step on an ant?).
Then it switches to parodying some arguments against the possibility of superintelligence, in order:
1. That intelligence is "irreducibly complex"; since we don't understand what makes people smart, we can't expect to make smarter machines. This, to me, is the weakest of the parodies in the article. We definitely understand size better than intelligence, and lack of fundamental understanding has absolutely impeded our ability to engineer things in the past. To get to the Moon, we had to understand gravitation at a much deeper level than "things fall down", or even than "things fall in parabolas". That said, total understanding of natural phenomena is not always necessary to replicate them: the Wright brothers didn't need to understood how exactly bird flight worked to build the first plane.
2. That intelligence is poorly defined ("human-level" AI is "meaningless"). "Largeness" is similarly vague, but an Airbus is still obviously bigger than a human. Vague goals can still be met.
3. That human intelligence is in some way "universal". I don't understand this argument enough to explain the parody. Can someone else chime in?
4. Psychological arguments against superintelligence, which says that the origin of belief-in and fear-of superintelligent machines is really evolutionary, not logical. This is a classic non-sequitur; if evolution predisposes humans to be afraid of wolves, it does not follow that humans have no reason to be afraid of wolves.
5. That the goal of AI research is to augment human intelligence, not supplant it. Another non-sequitur.
6. Arguments from the hard problem of consciousness, and philosophical questions like "can a machine ever be conscious?". Yet another non sequitur: even if you think a machine fundamentally can't be conscious, it doesn't follow that it can't be functionally superintelligent.
7. Bullshit arguments involving quantum mechanics, Godelian incompleteness, and other things with pop-cultural cachet that most people who invoke them don't have the technical chops to understand.
This is a reference to the principle of computational equivalence. All Turing-complete systems are capable of emulating each other. A human can execute any program a machine can (only slower), and derive benefit from the results even if he doesn't fully understand the process.
http://mathworld.wolfram.com/PrincipleofComputationalEquival...
I say the burden of proof here falls on those supporting artificial superintelligences. It's purely fictional at this point
From what I understand it's basically how using tools to solve a problem is still considered 'intelligent', so to measure intelligence you should consider what problems you can solve, not how. In particular creating a device to solve a problem would still count as solving the problem. Hence if humans can create it, it can't be more intelligent than humans.
The argument does make more sense for intelligence. Even if you don't agree with identifying 'problem solving skill' with 'intelligence' you still have the paradox that it is seemingly impossible to create something that can solve more problems than its creator.
For every machine of any size, a lot of effort goes into its creation. They also don't survive without external intervention. The larger the machine, the more effort that goes into their creation and maintenance (and the more vulnerable they are to minor failures).
Aircraft, as a singular example, have to be virtually torn apart inspected and repaired every couple hundred hours to ensure they won't fall apart during regular use. They are also susceptible to the loss of a single critical piece; the loss of which will cause them to fail, often spectacularly.
Big bridges require almost daily maintenance to keep them safe. Boats are incapable of the simple act of fueling themselves. Computers stop working properly due to cosmic radiation flipping bits. Knives get dull. Batteries run out. Radioactive isotope power generators decay. Plastic breaks down. Stone erodes.
We have, despite hundreds of thousands of years, barely been able create an object which can last for even a couple of hundred years without constant maintenance. We have yet to create a machine of any size which is capable of maintaining itself. Why would a computer program, itself reliant on one of our least reliable and fragile constructs, prove to be any different?
"The crows assert that a single crow could destroy the heavens. This is certainly true, but it proves nothing against the heavens, because heaven means precisely: the impossibility of crows"
I'm still not sure I am wrong.
The topic is somewhat similar, and the paper is interesting so far, but I am reserving my opinion for when I will be finished reading it.
Take that, Dave Chalmers!
It's telling us: we live in a material universe, intelligence is just a function of physical attributes. If you want to deny the possibility of machine intelligence, the burden of proof is on you because, denying that would be so ridiculous as denying supersized machines.
>2 It is worth clarifying that there are, of course, systems today that appear to exceed human size in narrow dimensions. Lamp posts are one example. The predictions that we are considering concern some more general notion of largeness.