= 27d567bc2d48d9f40e9ccc20ea370b15)
SarahC comments on Existential Risk and Public Relations - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (613)
I'm not sure what you refer to by "actual AI." There is a sub-field of academic computer science which calls itself "Artificial Intelligence," but it's not clear that this is anything more than a label, or that this field does anything more than use clever machine learning techniques to make computer programs accomplish things that once seemed to require intelligence (like playing chess, driving a car, etc.)
I'm not sure why it is a requirement that an organization concerned with the behavior of hypothetical future engineered minds would need to be in contact with these researchers.
Yes, the subfield of computer science is what I'm referring to.
I'm not sure that the difference between "clever machine learning techniques" and "minds" is as hard and fast as you make it. A machine that drives a car is doing one of the things a human mind does; it may, in some cases, do it through a process that's structurally similar to the way the human mind does it. It seems to me that machines that can do these simple cognitive tasks are the best source of evidence we have today about hypothetical future thinking machines.
I gave the wrong impression here. I actually think that machine learning might be a good framework for thinking about how parts of the brain work, and I am very interested in studying machine learning. But I am skeptical that more than a small minority of projects where machine learning techniques have been applied to solve some concrete problem have shed any light on how (human) intelligence works.
In other words, I largely agree with Ben Goertzel's assertion that there is a fundamental difference between "narrow AI" and AI research that might eventually lead to machines capable of cognition, but I'm not sure I have good evidence for this argument.
Although one should be very, very careful not to confuse the opinions of someone like Goertzel with those of the people (currently) at SIAI, I think it's fair to say that most of them (including, in particular, Eliezer) hold a view similar to this. And this is the location -- pretty much the only important one -- of my disagreement with those folks. (Or, rather, I should say my differing impression from those folks -- to make an important distinction brought to my attention by one of the folks in question, Anna Salamon.) Most of Eliezer's claims about the importance of FAI research seem obviously true to me (to the point where I marvel at the fuss that is regularly made about them), but the one that I have not quite been able to swallow is the notion that AGI is only decades away, as opposed to a century or two. And the reason is essentially disagreement on the above point.
At first glance this may seem puzzling, since, given how much more attention is given to narrow AI by researchers, you might think that someone who believes AGI is "fundamentally different" from narrow AI might be more pessimistic about the prospect of AGI coming soon than someone (like me) who is inclined to suspect that the difference is essentially quantitative. The explanation, however, is that (from what I can tell) the former belief leads Eliezer and others at SIAI to assign (relatively) large amounts of probability mass to the scenario of a small set of people having some "insight" which allows them to suddenly invent AGI in a basement. In other words, they tend to view AGI as something like an unsolved math problem, like those on the Clay Millennium list, whereas it seems to me like a daunting engineering task analogous to colonizing Mars (or maybe Pluto).
This -- much more than all the business about fragility of value and recursive self-improvement leading to hard takeoff, which frankly always struck me as pretty obvious, though maybe there is hindsight involved here -- is the area of Eliezer's belief map that, in my opinion, could really use more public, explicit justification.
I don't think this is a good analogy. The problem of colonizing Mars is concrete. You can make a TODO list; you can carve the larger problem up into subproblems like rockets, fuel supply, life support, and so on. Nobody knows how to do that for AI.
OK, but it could still end up being like colonizing Mars if at some point someone realizes how to do that. Maybe komponisto thinks that someone will probably carve AGI in to subproblems before it is solved.
Well, it seems we disagree. Honestly, I see the problem of AGI as the fairly concrete one of assembling an appropriate collection of thousands-to-millions of "narrow AI" subcomponents.
Perhaps another way to put it would be that I suspect the Kolmogorov complexity of any AGI is so high that it's unlikely that the source code could be stored in a small number of human brains (at least the way the latter currently work).
EDIT: When I say "I suspect" here, of course I mean "my impression is". I don't mean to imply that I don't think this thought has occurred to the people at SIAI (though it might be nice if they could explain why they disagree).
The portion of the genome coding for brain architecture is a lot smaller than Windows 7, bit-wise.
An oddly somewhat relevant article on the information needed for specifying the brain. It is a biologist tearing a strip out of kurzweil for suggesting that we'll be able reverse engineer the human brain in a decade by looking at the genome.
P.Z. is misreading a quote from a secondhand report. Kurzweil is not talking about reading out the genome and simulating the brain from that, but about using improvements in neuroimaging to inform input-output models of brain regions. The genome point is just an indicator of the limited number of component types involved, which helps to constrain estimates of difficulty.
Edit: Kurzweil has now replied, more or less along the lines above.
Kurzweil's analysis is simply wrong. Here's the gist of my refutation of it:
"So, who is right? Does the brain's design fit into the genome? - or not?
The detailed form of proteins arises from a combination of the nucleotide sequence that specifies them, the cytoplasmic environment in which gene expression takes place, and the laws of physics.
We can safely ignore the contribution of cytoplasmic inheritance - however, the contribution of the laws of physics is harder to discount. At first sight, it may seem simply absurd to argue that the laws of physics contain design information relating to the construction of the human brain. However there is a well-established mechanism by which physical law may do just that - an idea known as the anthropic principle. This argues that the universe we observe must necessarily permit the emergence of intelligent agents. If that involves a coding the design of the brains of intelligent agents into the laws of physics then: so be it. There are plenty of apparently-arbitrary constants in physics where such information could conceivably be encoded: the fine structure constant, the cosmological constant, Planck's constant - and so on.
At the moment, it is not even possible to bound the quantity of brain-design information so encoded. When we get machine intelligence, we will have an independent estimate of the complexity of the design required to produce an intelligent agent. Alternatively, when we know what the laws of physics are, we may be able to bound the quantity of information encoded by them. However, today neither option is available to us."
Wired really messed up the flow of the talk in that case. Is it based off a singularity summit talk?
I agree with your analysis, but I also understand where PZ is coming from. You write above that the portion of the genome coding for the brain is small. PZ replies that the small part of the genome you are referring to does not by itself explain the brain; you also need to understand the decoding algorithm - itself scattered through the whole genome and perhaps also the zygotic "epigenome". You might perhaps clarify that what you were talking about with "small portion of the genome" was the Kolmogorov complexity, so you were already including the decoding algorithm in your estimate.
The problem is, how do you get the point through to PZ and other biologists who come at the question from an evo-devo PoV? I think that someone ought to write a comment correcting PZ, but in order to do so, the commenter would have to speak the languages of three fields - neuroscience, evo-devo, and information-theory. And understand all three well enough to unpack the jargon to laymen without thereby loosing credibility with people who do know one or more of the three fields.
Obviously the genome alone doesn't build a brain. I wonder how many "bits" I should add on for the normal environment that's also required (in terms of how much additional complexity is needed to get the first artificial mind that can learn about the world given additional sensory-like inputs). Probably not too many.
Thanks, this is useful to know. Will revise beliefs accordingly.
What do you think you know and how do you think you know it? Let's say you have a thousand narrow AI subcomponents. (Millions = implausible due to genome size, as Carl Shulman points out.) Then what happens, besides "then a miracle occurs"?
What happens is that the machine has so many different abilities (playing chess and walking and making airline reservations and...) that its cumulative effect on its environment is comparable to a human's or greater; in contrast to the previous version with 900 components, which was only capable of responding to the environment on the level of a chess-playing, web-searching squirrel.
This view arises from what I understand about the "modular" nature of the human brain: we think we're a single entity that is "flexible enough" to think about lots of different things, but in reality our brains consist of a whole bunch of highly specialized "modules", each able to do some single specific thing.
Now, to head off the "Fly Q" objection, Iet me point out that I'm not at all suggesting that an AGI has to be designed like a human brain. Instead, I'm "arguing" (expressing my perception) that the human brain's general intelligence isn't a miracle: intelligence really is what inevitably happens when you string zillions of neurons together in response to some optimization pressure. And the "zillions" part is crucial.
(Whoever downvoted the grandparent was being needlessly harsh. Why in the world should I self-censor here? I'm just expressing my epistemic state, and I've even made it clear that I don't believe I have information that SIAI folks don't, or am being more rational than they are.)
If a thousand species in nature with a thousand different abilities were to cooperate, would they equal the capabilities of a human? If not, what else is missing?
Tough problem. My first reaction is 'yes', but I think that might be because we're assuming cooperation, which might be letting more in the door than you want.
Yes, if there were a sufficiently powerful optimization process controlling the form of their cooperation.
The brain has many different components with specializations, but the largest and human dominant portion, the cortex, is not really specialized at all in the way you outline.
The cortex is no more specialized than your hard drive.
Its composed of a single repeating structure and associated learning algorithm that appears to be universal. The functional specializations that appear in the adult brain arise due to topological wiring proximity to the relevant sensory and motor connections. The V1 region is not hard-wired to perform mathematically optimal gabor-like edge filters. It automatically evolves into this configuration because it is the optimal configuration for modelling the input data at that layer, and it evolves thus soley based on exposure to said input data from retinal ganglion cells.
You can think of cortical tissue as a biological 'neuronium'. It has a semi-magical emergent capacity to self-organize into an appropriate set of feature detectors based on what its wired to. more on this
All that being said, the inter-regional wiring itself is currently less understood and is probably more genetically predetermined.
There may be other approaches that are significantly simpler (that we haven't yet found, obviously). Assuming AGI happens, it will have been a race between the specific (type of) path you imagine, and every other alternative you didn't think of. In other words, you think you have an upper bound on how much time/expense it will take.
I'm not a member of SIAI but my reason for thinking that AGI is not just going to be like lots of narrow bits of AI stuck together is that I can see interesting systems that haven't been fully explored (due to difficulty of exploration). These types of systems might solve some of the open problems not addressed by narrow AI.
These are problems such as
Now I also doubt that these systems will develop quickly when people get around to investigating them. And they will have elements of traditional narrow AI in as well, but they will be changeable/adaptable parts of the system, not fixed sub-components. What I think needs is exploring is primarily changes in software life-cycles rather than a change in the nature of the software itself.
I don't think AGI in a few decades is very farfetched at all. There's a heckuvalot of neuroscience being done right now (the Society for Neuroscience has 40,000 members), and while it's probably true that much of that research is concerned most directly with mere biological "implementation details" and not with "underlying algorithms" of intelligence, it is difficult for me to imagine that there will still be no significant insights into the AGI problem after 3 or 4 more decades of this amount of neuroscience research.
Of course there will be significant insights into the AGI problem over the coming decades -- probably many of them. My point was that I don't see AGI as hard because of a lack of insights; I see it as hard because it will require vast amounts of "ordinary" intellectual labor.
I'm having trouble understanding how exactly you think the AGI problem is different from any really hard math problem. Take P != NP, for instance the attempted proof that's been making the rounds on various blogs. If you've skimmed any of the discussion you can see that even this attempted proof piggybacks on "vast amounts of 'ordinary' intellectual labor," largely consisting of mapping out various complexity classes and their properties and relations. There's probably been at least 30 years of complexity theory research required to make that proof attempt even possible.
I think you might be able to argue that even if we had an excellent theoretical model of an AGI, that the engineering effort required to actually implement it might be substantial and require several decades of work (e.g. Von Neumann architecture isn't suitable for AGI implementation, so a great deal of computer engineering has to be done).
If this is your position, I think you might have a point, but I still don't see how the effort is going to take 1 or 2 centuries. A century is a loooong time. A century ago humans barely had powered flight.
I think the following quote is illustrative of the problems facing the field:
-Marvin Minsky, quoted in "AI" by Daniel Crevier.
Some notes and interpretation of this comment:
I think this kind of observation justifies AI-timeframes on the order of centuries.
Edge detection is rather trivial. Visual recognition however is not, and there certainly are benchmarks and comparable results in that field. Have you browsed the recent pubs of Poggio et al at MIT vision lab? There is lots of recent progress, with results matching human levels for quick recognition tasks.
Also, vision is not a tiny part of intelligence. Its the single largest functional component of the cortex, by far. The cortex uses the same essential low-level optimization algorithm everywhere, so understanding vision at the detailed level is a good step towards understanding the whole thing.
And finally and most relevant for AGI, the higher visual regions also give us the capacity for visualization and are critical for higher creative intelligence. Literally all scientific discovery and progress depends on this system.
"visualization is the key to enlightenment" and all that
the visual system
It's only trivial if you define an "edge" in a trivial way, e.g. as a set of points where the intensity gradient is greater than a certain threshold. This kind of definition has little use: given a picture of a tree trunk, this definition will indicate many edges corresponding to the ridges and corrugations of the bark, and will not highlight the meaningful edge between the trunk and the background.
I don't believe that there is much real progress recently in vision. I think the state of the art is well illustrated by the "racist" HP web camera that detects white faces but not black faces.
I actually agree with you about this, but I think most people on LW would disagree.
By no means do I want to downplay the difficulty of P vs NP; all the same, I think we have different meanings of "vast" in mind.
The way I think about it is: think of all the intermediate levels of technological development that exist between what we have now and outright Singularity. I would only be half-joking if I said that we ought to have flying cars before we have AGI. There are of course more important examples of technologies that seem easier than AGI, but which themselves seem decades away. Repair of spinal cord injuries; artificial vision; useful quantum computers (or an understanding of their impossibility); cures for the numerous cancers; revival of cryonics patients; weather control. (Some of these, such as vision, are arguably sub-problems of AGI: problems that would have to be solved in the course of solving AGI.)
Actually, think of math problems if you like. Surely there are conjectures in existence now -- probably some of them already famous -- that will take mathematicians more than a century from now to prove (assuming no Singularity or intelligence enhancement before then). Is AGI significantly easier than the hardest math problems around now? This isn't my impression -- indeed, it looks to me more analogous to problems that are considered "hopeless", like the "problem" of classifying all groups, say.
I hate to go all existence proofy on you, but we have an existence proof of a general intelligence - accidentally sneezed out by natural selection, no less, which has severe trouble building freely rotating wheels - and no existence proof of a proof of P != NP. I don't know much about the field, but from what I've heard, I wouldn't be too surprised if proving P != NP is harder than building FAI for the unaided human mind. I wonder if Scott Aaronson would agree with me on that, even though neither of us understand the other's field? (I just wrote him an email and asked, actually; and this time remembered not to say my opinion before asking for his.)
Scott says that he thinks P != NP is easier / likely to come first.
Well, I for one strongly hope that we resolve whether P = NP before we have AI since a large part of my estimate for the probability of AI being able to go FOOM is based on how much of the complexity hierarchy collapses. If there's heavy collapse, AI going FOOM Is much more plausible.
Obviously not. That would be a proof of P != NP.
As for existence proof of a general intelligence, that doesn't prove anything about how difficult it is, for anthropic reasons. For all we know 10^20 evolutions each in 10^50 universes that would in principle allow intelligent life might on average result in 1 general intelligence actually evolving.
Well actually, after thinking about it, I'm not sure I would either. There is something special about P vs NP, from what I understand, and I didn't even mean to imply otherwise above; I was only disputing the idea that "vast amounts" of work had already gone into the problem, for my definition of "vast".
Scott Aaronson's view on this doesn't move my opinion much (despite his large contribution to my beliefs about P vs NP), since I think he overestimates the difficulty of AGI (see your Bloggingheads diavlog with him).
Awesome! Be sure to let us know what he thinks. Sounds unbelievable to me though, but what do I know.
Why is AGI a math problem? What is abstract about it?
We don't need math proofs to know if AGI is possible. It is, the brain is living proof.
We don't need math proofs to know how to build AGI - we can reverse engineer the brain.
This is a good part of the guts of it. That bit of it is a math problem:
http://timtyler.org/sequence_prediction/
There may be a few clues in there - but engineers are likely to get to the goal looong before the emulators arrive - and engineers are math-friendly.
...but you don't really know - right?
You can't say with much confidence that there's no AIXI-shaped magic bullet.
That's right; I'm not an expert in AI. Hence I am describing my impressions, not my fully Aumannized Bayesian beliefs.
AIXI-shaped magic bullet?
AIXI's contribution is more philosophical than practical. I find a depressing over-emphasis of bayesian probability theory here as the 'math' of choice vs computational complexity theory, which is the proper domain.
The most likely outcome of a math breakthrough will be some rough lower and or upper bounds on the shape of the intelligence over space/time complexity function. And right now the most likely bet seems to be that the brain is pretty well optimized at the circuit level, and that the best we can do is reverse engineer it.
EY and the math folk here reach a very different conclusion, but I have yet to find his well considered justification. I suspect that the major reason the mainstream AI community doesn't subscribe to SIAI's math magic bullet theory is that they hold the same position outline above: ie that when we get the math theorems, all they will show is what we already suspect: human level intelligence requires X memory bits and Y bit ops/second, where X and Y are roughly close to brain levels.
This, if true, kills the entirety of the software recursive self-improvement theory. The best that software can do is approach the theoretical optimum complexity class for the problem, and then after that point all one can do is fix it into hardware for a further large constant gain.
I explore this a little more here
The article linked to in the parent is entitled:
"Created in the Likeness of the Human Mind: Why Strong AI will necessarily be like us"
Good quality general-purpose data-compression would "break the back" of the task of buliding synthetic intelligent agents - and that's a "simple" math problem - as I explain on: http://timtyler.org/sequence_prediction/
At least it can be stated very concisely. Solutions so far haven't been very simple - but the brain's architecture offers considerable hope for a relatively simple solution.
That seems like crazy talk to me. The brain is not optimal - not its hardware or software - and not by a looooong way! Computers have already steam-rollered its memory and arithmetic -units - and that happened before we even had nanotechonolgy computing components. The rest of the brain seems likely to follow.
Note that allowing for a possibility of sudden breakthrough is also an antiprediction, not a claim for a particular way things are. You can't know that no such thing is possible, without having understanding of the solution already at hand, hence you must accept the risk. It's also possible that it'll take a long time.
I'm reading through and catching up on this thread, and rather strongly agreed with your statement:
However, pondering it again, I realize there is an epistemological spectrum ranging from math on the one side to engineering on the other. Key insights into new algorithms can undoubtedly speed up progress, and such new insights often can be expressed as pure math, but at the end of the day it is a grand engineering (or reverse engineering) challenge.
However, I'm somewhat taken aback when you say, "the notion that AGI is only decades away, as opposed to a century or two."
A century or two?
One obvious piece of evidence is that many forms of narrow learning are mathematically incapable of doing much. There are for example a whole host of theorems about what different classes of neural networks can actually recognize, and the results aren't very impressive. Similarly, support vector machine's have a lot of trouble learning anything that isn't a very simple statistical model, and even then humans need to decide which stats are relevant. Other linear classifiers run into similar problems.
I work in this field, and was under approximately the opposite impression; that voice and visual recognition are rapidly approaching human levels. If I'm wrong and there are sharp limits, I'd like to know. Thanks!
Machine intelligence has surpassed "human level" in a number of narrow domains. Already, humans can't manipulate enough data to do anything remotely like a search engine or a stockbot can do.
The claim seems to be that in narrow domains there are often domain-specific "tricks" - that wind up not having much to do with general intelligence - e.g. see chess and go. This seems true - but narrow projects often broaden out. Search engines and stockbots really need to read and understand the web. The pressure to develop general intelligence in those domains seems pretty strong.
Those who make a big deal about the distinction between their projects and "mere" expert systems are probably mostly trying to market their projects before they are really experts at anything.
One of my videos discusses the issue of whether the path to superintelligent machines will be "broad" or "narrow":
http://alife.co.uk/essays/on_general_machine_intelligence_strategies/
Thanks, it always is good to actually have input from people who work in a given field. So please correct me if I'm wrong but I'm under the impression that
1) neutral networks cannot in general detect connected components unless the network has some form of recursion. 2) No one knows how to make a neural network with recursion learn in any effective, marginally predictable fashion.
This is the sort of thing I was thinking of. Am I wrong about 1 or 2?
Not sure what you mean about by 1), but certainly, recurrent neural nets are more powerful. 2) is no longer true; see for example the GeneRec algorithm. It does something much like backpropagation, but with no derivatives explicitly calculated, there's no concern with recurrent loops.
On the whole, neural net research has slowed dramatically based on the common view you've expressed; but progress continues apace, and they are not far behind cutting edge vision and speech processing algorithms, while working much more like the brain does.
Thanks. GeneRec sounds very interesting. Will take a look. Regarding 1, I was thinking of something like the theorems in chapter 9 in Perceptrons which shows that there are strong limits on what topological features of input a non-recursive neural net can recognize.