Transcending Complacency on Superintelligent Machines

http://www.huffingtonpost.com/stephen-hawking/artificial-intelligence_b_5174265.html

by hedges • 12 years ago

43 34 12 years ago

swombat• 12 years ago

Obviously agreeing with the points, but this seems to be more an awareness piece than an actionable one.

cliveowen• 12 years ago

Am I the only one who's skeptical about the feasibility of AI? As I see it there are two ways to think about AI: first there's the kind of AI that arises from software emulating parts of the human brain based on our current understanding of its inner functioning and produces human-like intelligence, so even if the mechanisms are different from those actually employed by the brain the output is similar in response and depth of reasoning; then there's the AI that stems from creating an artificial brain by reverse-engineering the human brain, but we are an awfully long way from doing that, mostly because we can't expect to unravel in a few decades what evolution has spent millennia perfecting.

It looks to me, a layman, that the only approach that holds any water is the first one. But then again, it mostly looks like people are implementing software based on a flawed understanding of cognitive functions and basically hoping that something magic happens. How can a scattershot approach like this ever produce anything even remotely resembling human intelligence?

_random_• 12 years ago

"mostly because we can't expect to unravel in a few decades what evolution has spent millennia perfecting" - that's not the kind of thinking that got us driving cars, flying planes, coding computers and altering genes. Evolution doesn't have a brain.

mrqwerty• 12 years ago

Evolution is a brain.

Gobitron5• 12 years ago

I'm extremely skeptical of this, even though I spend a good deal of my time trying to automate machine 2 machine communication (which gets me thinking about this a lot). I also think that there's a good amount of hubris in these scientists believing that the brain is something that can be emulated by computing power of any kind...are we sure it's an apt analogy (brain as computer)?

codyb• 12 years ago

It's the holy grail of the field.

gamegoblin• 12 years ago

On an abstract level, the brain is just a complex circuit. Neurons take inputs as electric signals and send outputs as electric signals.

Neurons are bit more general than our typical logic gates, but at a high level they have a lot in common.

solipsism3• 12 years ago

Where is the hubris in attempting to do something without knowing whether it's possible? What is it about AI that seems to make people defensive? Is it the idea that human intelligence might not be as special as we think it is?

sebastialonso3• 12 years ago

Why not? I keep seeing these discussions pointing out the futility of thinking about the brain as a computer, but still don't see powerful arguments to back that postulate up, other than the 'brain too hard, computer too easy, so brain no computer'.

Brain as a computer, in my opinion should be the default state for this discussions. Why? Consider the old and tired brain-made-of-matter argument. There's no reason to think there's something magical or supernatural inside the brain, so treat it as an organized collection of atoms doing cool stuff. The default state cannot be magic, it has to be something that can be disproved or ruled out.

Some parts of it seem to work, as fas as we know, in a (suspiciously) algorithmic way, or in other words, a highly abstract step-by-step chain of actions can be identify for a given part of the brain.

Why not start with the crazy assumption that the whole brain acts as a computer (the theoretical concept), and then identify which parts of it fail the analogy? The key part here is the word 'fail': it should not mean 'too complex for any computer we have built' nor 'we don't know any algorithm that does that', it should mean that there are parts that inherently cannot be modelled, under any circumstances, like the definition of algorithm. If some part is discovered not to hold the analogy, you should just then question if the analogy is question is apt or not.

pygy_• 12 years ago

You don't need to emulate brains to create intelligence anymore than you have to imitate birds to build airplanes.

krapp2• 12 years ago

Isn't that an open question?

Also, airplanes do imitate birds in a few fundamental ways.

bermanoid• 12 years ago

Evolution did just fine, and it's far more scattershot than our current efforts at AI.

You're also crucially missing the possibility that someone comes up with an intelligent algorithm that doesn't mirror the human brain in much detail, but still manages to outperform it. Think of flight: inventing flapping machines didn't turn out to be very useful, but we figured out a workaround that was far more efficient. The most interesting (terrifying?) AI research is along these lines.

dimatura2• 12 years ago

Is it feasible? Yes, absolutely. Are we anywhere near to 'human-level' AI? No, not by a long shot. How long will it take? Will we ever make it? Very hard to know.

tluyben26• 12 years ago

Well... That definition keeps changing. For passed definitions we passed human-level a few times. And for some tasks we did or will pass it soon. So what is your definition of 'human-level' intelligence? Don't copy wikipedia; make one up in strict language (don't say consciousness please). For what a lot people make/made up we passed.

solipsism3• 12 years ago

mostly because we can't expect to unravel in a few decades what evolution has spent millennia perfecting.

As others have suggested, this is demonstrably wrong, trivially so.

EzraVinh2• 12 years ago

Actually, there are some areas where evolution still has an advantage. A biologist was explaining to me how artificial enzymes (man made proteins) can at best speed up a chemical reaction by 10^3 - 10^6, while natural enzymes typically speed them up by 10^9 - 10^12. Might have got those number wrong, but his point was that natural enzymes beat artificial ones by a wide margin.

hyp0• 12 years ago

You're being a bit parochial, specificially anthropocentric. Why a human brain? If you consider it's possible that other intelligences exist in the universe, they probably won't have human brains... Or, consider other animals on earth: given several million years and the right conditions, do you consider it possible that intelligence (at our level) could evolve again - and I don't just mean apes, dolphins and elephants. Especially fascinating in this connection are parrots, crows and other corvids, which are obviously quite intelligent, but have a completely different neuroanatomy of the higher centers from mammals.

Are really we so special? Apart from being first (that we know of)? The history of science is a history of anthropocentrism iconoclasm!

Of course, copying what we know works seems like a reasonable starting point.

Secondly, who said it has to be done within a few decades? If it is done within centuries or millennia, it is still done. "Infeasible" would mean it can't be done, ever.

Personally, I tend towards 50 years off... as it has been for a long time (and probably will continue to be) <- this is a joke. I'm saying it's a looong way off.

EzraVinh2• 12 years ago

I agree with your assessment on those two avenues. One thing that throws people off is that the AI/Machine Learning community is constantly selling their models as "human like" when the models are really only inspired by the human brain by extremely loose analogies.

There is another approach, what I would call the Airplane approach, since it is to the brain what airplanes are to birds. That is, to base machine intelligence on a new kind of mathematical logic that hasn't been invented yet.

TheIronYuppie• 12 years ago

I think that one thing that people are missing when they think AI is not a threat is there does not have to be a singular AI for every problem.

Chess computers are better than humans, but I wouldn't trust them to manage the electricity grid. What if there was an equivalent quality of computer specialized for every significant area of society - electricity grid, packet routing, high speed trading, etc etc.

todd8• 12 years ago

I'm no expert in this area so I welcome corrections to my rough numbers below.

    Approximate number of human neurons:    1.0e11
    Approximate number synapses in a human: 1.0e14

These are big numbers, but not impossibly big numbers. There are different kinds of neurons, and signals on synapses are not simply binary. However, even with these complications, the hardware needed to reach these scales isn't hard to imagine.

    Transistors in XBox One:                1.0e09

Brains are biological computers so they suffer from very slow switching speeds at the neural level. Neurons run in parallel, but they are not fast:

    Approximate neural switching speed:     1.0e03/sec

Even if all of the synapeses could sustain this rate in parallel (they can't) and even if all of the brain was 100% occupied with solving a single task (it isn't) this would mean that we absolutely can't compute faster than:

    Speed * synapses (brain ops per second): 1.0e17/sec

For comparison, the fastest bitcoin hardware I see is advertised to operate at the following speed:

    Minerscube 15 (hashes per second):      1.5e12/sec

And a regular GPU is capable of simple instructions that run at the following speed:

    AMD Radeon HD 6990 32-bit instructions: 2.6e12/sec

From this we can see that hardware is catching up with the raw computing ability of the human brain. Now consider the problem of programming a brain. It isn't necessary to program every synapse. The brain learns, and essentially programs itself. To see why this is true consider the programming that we are born with:

    Bits of information in human genome:    1.0e10

This is far less than the number of synapses that we have. Therefore, the brain must program itself, somehow.

Now, to address the argument about evolution taking millions of years. First, we can evolve programs much faster than nature can evolve humans. There have been, perhaps, 100 million generations of humans. Even if it takes six seconds of computing time to run an evolutionary computation for a single generation it will take no more than 20 years to run over 100 million generations.

Brute force evolution isn't the only way to build strong AI. A program can exhibit behavior that we don't anticipate. I've written simple programs that beat me easily at games such as Othello or Freecell.

Finally, once machines get smart enough to design other machines there may be a rapid acceleration of progress in this area as we employ them in designing subsequent generations.

I feel that strong AI may pose a significant risk to humans; consequently, we should proceed with caution. Here is a thought experiment. If a chimpanzee could be taught to drive, would you trust it to pick your kids up from school? What sort of value judgements would it make in the case of an impending emergency? Would you let an elephant baby sit for you? Even if was much "smarter" than a normal elephant?

Strong AI will not be like us. It will learn and develop without a human body, and it will not interact with the world and society as we do and may end up being very foreign to us. Will it be sociopathic? Or will it be like whales, intelligent, but mysterious, perhaps spending all its time singing AI songs to other AIs.

Transcending Complacency on Superintelligent Machines

Related Stories