People who ask questions such as this often don't consider that it remains eminently possible that AGI is an impossibility for us to build. Also remember that anything an AI can do in the future a human + an AI can probably do better. Right now at least they're just tools we use and will remain so for the foreseeable future.
https://en.wikipedia.org/wiki/The_Emperor's_New_Mind
NB I didn't find this argument particularly strong, but it's a long time since I read it. The brain would have to be doing something spectacularly strange for it to be impossible for us to emulate in one way or another.
More generally, I don't think invoking quantum effects is an especially strong argument for impossibility in the general case. We already use quantum effects to some extent and make progress on quantum computation. It may be evidence of greater difficulty, but not of impossibility. Unless there is a strong case for some particular quantum effect being unharness-able.
Not that quantum mechanisms in brains are likely, so probably not an issue.
The issue is that we do not have proof that general AI is possible. We conjecture that it is based on the assumption that our brains could be simulated by enough artificial neurons, (or some other method) but we don't know that for a fact.
And so far we can't even fully simulate a mouse brain, let alone a human brain. We can make something that imitates aspects of mouse behavior, but we can't actually simulate a mouse brain.
All the hype around General AI being "just over the horizon" right now is just like it was back in the 80s (or the 60s.) It's a bit over the top. If you look at it from a critical perspective there isn't a lot of solid reason to be so bullish.
It's a very long way off.
Based on everything else we know about the world now, we can "just" simulate a brain. "Just" may mean "after another 200 years of work", of course. The alternative is "the brain is magical somehow", which is not a strong argument on its own – we don't often end up finding magic to be the answer.
Where do you see hype about General AI being just over the horizon? Most popular discussion I read is about two things: a) the replacement of human work by increasingly-improved specific AI; b) the possibility of rapid advancement of general AI once invented. I haven't seen anyone say b) is imminent – only that it is likely to be fast/dangerous when it happens.
Right, we think we can. But we don't know. I would argue that as we get closer to doing it, it will become a fact that we can (or we will find that we can't.) The fact that we still don't know for sure that we can (or cannot) is an indicator or a symptom of how far away we are from actually doing it.
Is the argument that we're still just hitting the low-hanging fruit of the technological age, so have false confidence in all the things we've equalled or beaten?
I guess what I'm trying to get at is that most of the hardest things may not be logically possible (c.f. incompleteness theorem) or may not be physically possible (c.f. time travel). Having such clear physical possibility (and at multiple levels of complexity/advancement that only arose through selection and mutation) seems like strong evidence we will be able to do it eventually. A claim of impossibility is not illogical, but requires a stronger argument than "there's always a small chance".
In order to accurately simulate a particular brain, we'd need the technology to (i) accurately read all the information encoded in a particular brain, which everything we know about the brain suggests depends on getting a snapshot of the state of 85 billion neurons - and chemical interactions of those neurons' constituent proteins - at a particular millisecond in time. (ii) manipulate matter (or write matter-simulating code) such that we duplicate the state of that brain at that time, and probably the associated body too, with sufficient accuracy for the biochemical reactions (or digital emulation of the biochemical reactions) continue to produce a patterns resembles the thought processes of the emulated brain in their outputs.
I don't think we need to invoke "magic" or vitalism or quantum brains to suggest that replicators capable of reproducing a particular state of a mind-bogglingly complex set of biochemical processes without significant information loss may not be a practical possibility in future. If anything, that sounds harder than achieving AGI or something approaching it via other means.
As for the hype about general AI, we're responding to a thread musing on whether we'll get "close" to achieving it this year.
If "all the hype" is a single HN question (not even an assertion) this counts as "over the top", then what would be an appropriate amount of discussion other than none at all?
Someone who probably isn't wholly uninformed asking if we're getting close to AGI this year is obviously a symptom rather than the cause of the hype. I mean, if we're working on Ray Kurzweil's Singularity is Near timescale, I'll be able to buy computational power equivalent to the human brain for $999 in three years' time. Still, Peter Diamindis did update that a couple of years ago by pushing that deadline out to 2025, so I guess at least I'll still be smarter than my next bottom-spec Macbook Air.
[1]C elegans is pretty trivial
It's not, though. It comes up over and over here, and often it's a link to an external article rather than an "Ask HN".
I also wish all of the sugar beverage hype would stop (https://hn.algolia.com/?query=coca%20cola&sort=byPopularity&...).
I don't really buy that argument (even though it's mine, I've not seen anyone else argue this). We design systems so complex and built from learning algorithms and neural nets such that ultimately we don't actually know precisely how they are doing what they do. We could brute force it by evolving a solution maybe. But it's at least plausible that we won't be able to engineer and design an AI architecturally because we're not up to the task.
Its just lots of that stuff from the waitbutwhy article really rankles me: e.g. having an x-axis for "human advancement" and then only valuing recent developments while under valuing things like writing or the printing press or treating intelligence as an integer value where an AGI can keep multiplying its intelligence and become completely unassailable by humans. There's a lot of bollocks intertwined with people that hope for AGI and it upsets me. I feel like it devalues and undermines the current progress that is being made.
My own view is that it's highly likely that we'll have something like AGI - but we might also find that it's not actually that useful, or really necessary, and anyway I suspect it will take a long time to get there (I'd guess 100-200 years).
I'm particularly fond of the Back to the Future example of how rapidly and unpredictably technological progress was accelerating, which completely ignored the fact that Back to the Future had a sequel whose tongue-in-cheek version of 2015 based entirely on 1980s culture is pretty much dead on, bar the hovering/flying stuff we haven't got around to inventing yet.
In other words, he's lying with a graph in order to try to make his argument have more emotional impact. And I have a policy of taking lying as a serious disqualification of people who are trying to persuade me...
I 100% agree that AGI is probably a long ways off barring some surprise (not that such a surprise is impossible, but definitely not expected). That said, the fact that the human brain works means that brains can be made. Unless you think there's something literally magical going on in the brain, then the brain is just a physical thing, and other physical things with similar behavior can be constructed. Maybe we won't be able to do it with current computer hardware, but some day we'll manage it.
I've never heard an argument that AGI was impossible that didn't basically boil down to "No, we're special, and you can't prove otherwise." Not a very convincing argument.
More the opposite. What makes you assume we're smart enough to fully understand how it is that we work and reproduce it non-biologically? My main issue is around some people treating it as an inevitability as opposed to a possibility. I mean I'm all for humans performing crazy engineering feats but we have to ask ourselves if some feats are beyond us. Could we build a star? Would we build a star? Will we build one 2017?
You get me?
Simply running basic neural network is slow work, even with the best supercomputers we have today. Throw in the fact that a neural network is a super simplistic model of how we currently believe synapses work, and the scale of the simulation problem gets hard, fast.
That's an interesting way to put it! I tend to agree that simulating a human brain will need more than the computational resources to train a neural network with 1 billion neurons (if those resources will ever be available, which is questionable). That's not to say we won't find a way around that problem though. Of course, all the worry about a dangerous AGI hinges on this premise, and how we get there is at best unclear at the moment.
Absolutely nothing. I believe we'll eventually make artificial intelligence, and I doubt we'll really understand it when we make it. Human intelligence evolved naturally as a result of a statistical process. When we make artificial intelligence it's totally likely that it will be the result of some human controlled statistical process that looks nothing like traditional software.
> Could we build a star?
Almost certainly eventually.
> Would we build a star?
Probably not. But building a star would require vast resources, and I'm not sure what we'd get out of it. There are lots of stars. Artificial intelligence really isn't comparable on this front.
1960s Herbert Simmons predicts "Machines will be capable, within 20 years, of doing any work a man can do."
1993 - Vernor Vinge predicts super-intelligent AIs 'within 30 years'.
2011 ray Kurzweil predicts the singularity (enabled by super-intelligent AIs) will occur by 2045, 34 years after the prediction was made.
So the distance into the future before we achieve strong AI and hence the singularity is, according to it's most optimistic proponents, receding by more than 1 year per year.
I am not in any way denying the achievability of strong AI. I do believe it will happen. I just don't think we currently have any idea how or when. If pushed to it, I'd say probably more than another 100 years from now but I don't know how much more.
This is different from, say, AlphaGo playing against itself to train its neural network - we want AI 1.0 to write AI 2.0, not just tweak some coefficients in 1.0.
At the moment all automatically generated code is less complex than source code of code generator itself. There can be more of it in terms of lines of code, but it's usually pretty repetitive.
This doesn't even begin to get into the core of AGI, which is the 'thinking' component. Given this amazing mass of data, how do we then make the machine work towards it's goals? Is this just a neural network? Is it a billion neural networks? Too many variables to tell.
And even then, if every action it takes is a reaction to the environment, does it then not have freewill? Do we have freewill? Is 'consciousness' somehow the key to freewill?
But anyway if you listen to Musk or Hawking, doomsday AI is just round the corner.
Ultimately I think there are far more pressing issues in AI around ethics, bias, or security. It's great that there are philosophers who can sit around and worry about a possible distant future (and I think as thought exercises they're fascinating), but it's not what most people in the field should be concerned with.
From the site:
"minimum bet period is two years"
I'd extend it much more because the real odds of AGI are closer to 1:100, but any longer time scale would be a bad return on my money.
People who are really working on the ideas discussed (as opposed to just talking about them and positioning themselves as "futurists") are accountable in a much more profound way, in that they build their entire lives around working toward some goal, with all the risk and opportunity cost that entails. The long bets participants are just playing around, not real players. They may have a little skin in the game, but it's faux skin.
[0] http://karpathy.github.io/2012/10/22/state-of-computer-visio...
Many responses in this thread are along the lines of "computers will never be as smart as Carl Sagan", ignoring that most intelligence on this planet could never dream of being half as smart as the genius we're using to define AI. Let's start by getting a computer as smart as a border collie, one of the more intelligent dogs.
Saying a computer isn't smart because it can't laugh at a highly complex visual joke is entirely the wrong way to define AI, and it's no surprise we haven't achieved it. It took humans millions of years to get this smart, and we've only been working on AI for a very, very small amount of time.
I think that the main obstacle is language - language is terribly ambiguous, and its very difficult to deal with these ambiguities in a program.
Hofstadter [1] says that the core of thinking are analogies and that many of the allusions in language can be thought of as analogies, however this does not seem to be the main focus of inquiry right now.
[1] https://www.amazon.com/Surfaces-Essences-Analogy-Fuel-Thinki... (my review & summary is here http://mosermichael.github.io/cstuff/all/blogg/2013/10/15/po... )
I don't think that means they are close to General AI though.
I never get this argument. Sure, maybe it is from the point of the computer, but we humans use it just fine.
Do you know how often I'm driving while my wife is navigating and I ask "turn left here?" and she says "right"? Now, is she saying "correct, you should turn left" or is she saying "don't turn left, turn right"?
Outside of a dog, a book is man’s best friend. Inside of a dog, it’s too dark to read. –Groucho Marx
I haven’t slept for ten days, because that would be too long. –Mitch Hedberg
Yes, but we laugh because we understand there is an ambiguity, not because we don't see it.
> How many petabytes on the Internet are wasted with comments correcting someone's use of language?
People are pedantic on the net, plus on the world wide web not everyone is going to be an English first speaker. Spoken conversations don't have people correcting your grammar every 5 seconds.
> wife is navigating and I ask "turn left here?" and she says "right"
Well in this case she is being deliberately ambiguous. So you either tell by her tone inflection, or rely on previous memory. And of course you have the ability to ask her.
> Outside of a dog, a book is man’s best friend. Inside of a dog, it’s too dark to read. –Groucho Marx
> I haven’t slept for ten days, because that would be too long. –Mitch Hedberg
Again, I see and understand the ambiguity. I'm not sitting here dumbfounded; I 'get' that they turned the words back on themselves to mean something else.
Will we get close in 2017? No. Not if my pet theory is right, and not if it's wrong.
https://en.wikipedia.org/wiki/Artificial_general_intelligenc...
What then?
General AI will have to wait until after that.
That isn't to say the resources don't exist to create AGI. It's possible they were available a long time ago. If you were to ask some omnipotent future superintelligence for a way humans could have bootstrapped AGI in the year 2005 using the available technology of the day, it could probably come up with an answer. Maybe even further back than that, or maybe even present day wouldn't suffice—who knows.
Trying to emulate biological architectures on silicon can be grossly inefficient, and may actually be harder from a design perspective. It is the attempt to formalize and adapt something created by an optimization process that spanned millions of years, a process that had zero regard for how easy its creation would be to understand or otherwise reverse engineer.
At the same time, algorithms vastly more efficient than the human brain's remain a possibility. They need not include the large amounts of evolutionary baggage that humans have.
Approaching AGI as a raw optimization problem may yield better results. However, not formally specifying or understanding the underlying mechanisms is a massive safety issue in the long run.
By the same token, ditching silicon entirely may be a vastly quicker path. Throwing ethics out the window and experimenting with large quantities of lab-grown neural tissue might be one way. Creating a synthetic biological computing substrate another. It's not hard to imagine something like copying human neural tissue's design, but using materials capable of latencies an order of magnitude lower, or significantly higher degrees of interconnectivity.
Looking at the problem from the perspective of strictly space, it's funny to think that we're unable to recreate the functionality of some tissue contained within a space that's less than one cubic foot-even though we have seemingly endless acres of computing power to do it with—that's excluding the brains of the thousands of scientists and engineers working on AI. Even if you stacked up just the microprocessors in question, they would occupy a cubic volume far, far greater than a single human brain—each containing billions of transistors, and each operating at latencies far lower than the brain. Despite all this, the human brain requires far lower amounts of energy.
The reason we don't have AGI yet is that it simply takes a lot of time and effort to invent, regardless if it's ultimately possible with today's technology. Of course, as other commenters have suggested, ruling out the possibility that the human brain somehow has seemingly magical quantum properties that render its recreation an impossibility (on silicon at least) may be unwise.
The term "Artificial Intelligence" is a contradiction - intelligence can NOT be artificial. Intelligence is the ability of a being to get what it wants. It is always organic, as it originates in desire.
Just stop calling it "Artificial Intelligence" and enjoy the wonderful progress that we are making towards getting our machines to help us achieve what we want.
(To be clear, I'm not saying stop calling it "artificial". I'm saying stop calling it "intelligence", because it is not, and never will be. Using the word "intelligence" in the context of machine automation sets entirely unreasonable expectations and inhibits progress. )
https://www.merriam-webster.com/dictionary/intelligence