Artificial Addition (2007)
lesswrong.com
lesswrong.com
It would be nice to have an updated version because the option 'we just need more computing power' (and training data) is now working pretty well for many machine learning domains (i.e. deeplearning).
It seems to me, (from a notably uneducated in AI perspective), that when dealing something deterministic like computing, general AI either works or it doesn't. You can't "do" something that's not conceptually solved in this domain, because what you are really doing is something else, albeit something that is very much like the goal.
Even Go was "conceptually solved" decades ago- it's "only" iterating over a search tree; we were just waiting for the hardware and tree-pruning algorithms to catch up. Humor and other human tasks can be more precisely defined but still don't even have theoretical algorithms to solve them.
I disagree that we can make computers do humor. Kobeya linked to a paper that shows unsupervised joke generation, but it's a stretch to call this humor. I haven't explored beyond what's in that article, but the examples of jokes produced do little to convince me that this program is doing humor:
I like my relationships like I like my source, open
I like my coffee like I like my war, cold
I like my boys like I like my sectors, bad
Show me an AI that can produce original joke forms and I'll reevaluate.
Go was not conceptually solved decades ago if it required the development of new tree-pruning algorithms to work. Improving an algorithm is conceptual innovation. If it could have worked with only a few orders of magnitude difference in performance, or with only marginally inferior results without new algorithms then maybe it was conceptually solved decades ago, else no.
Visual processing also comes to mind. Computers can "see", but it's very difficult to understand the many rules humans use to parse their environment, thus computers are not yet at human level of visual processing.
Makes me wonder if those things aren't "solved" by our brains, but we just don't quite know how. Another thing worth noting is that I know people who have no sense of humor. They simply do not understand jokes or sarcasm.
http://homepages.inf.ed.ac.uk/s0894589/petrovic13unsupervise...
The Wright brothers basically mimicked nature and tested till they got it right. They didn't use Bernoulli's equation to create a NACA airfoil and L/D max calculations. [2].
To that end, in one of my undergraduate Aeronautics classes, when we were doing a bunch of Bernoulli equations someone calculated that based on that approach helicopters couldn't actually fly. To which our instructor said, "Oh yea, helicopters fly by magic."
[1]https://secretofflight.wordpress.com/incorrect-theories/ [2]https://wright.nasa.gov/overview.htm
Creating aerodynamics simulation is creating an aerodynamic system. Creating a general intelligence is creating an intelligent system. A better analog to creating a helicopter using incomplete understanding of aerodynamics is something like creating a system that can discriminate between what is and isn't a general intelligence. We don't have such a system, but it at least doesn't seem like the sort of thing that would require a full understanding of intelligence.
Maybe automatic hyper-parameter optimization of adversarial networks is all that it takes for general intelligence. If you look up the conference proceedings of NIPS or AAAI you'll see major progress being made on basically every type of learning and memory you can find in a psychological textbook. Of course these are controlled experiments being done, in the technical sense, and not agglomerative. But what's stopping us from combining a bunch of these algorithms together, or better yet a base framework that generically supports them all, and making it self-reflective? Just computing power and available data, really.
So no, it's not at all clear that this is a problem that is "not conceptually solved." People keep saying that whenever the subject comes up, but the reality on the ground is otherwise.
These people wouldn't happen to be domain experts...?
I won't say that it's inconceivable that a sufficiently large and powerful neural net, with a simple structure based on simple rules, given a very comprehensive set of training data, could acheive a level of general intelligence.
But we don't have evidence of intelligence arising that way in the past. What we have is the human brain, an incredibly complex structure resembling a neural net with simple rules for individual node (maybe?). It's the result of evolutionary processes, which means that we should expect it to have extremely hard-to-comprehend properties (example [0]). We've seen evolutionary processes produce intelligence, but not learning processes. Again, that's not to say that evolutionary processes are necessary, just that we can be sure that they can lead to intelligence, and that we can't be sure for other methods.
[0] https://www.damninteresting.com/on-the-origin-of-circuits/
Sure:
https://www.amazon.com/Engineering-General-Intelligence-Part...
> Maybe what you mean to say is that once we have the computing power and data collection technology required, so that researches can experiment, the unsolved conceptual problems will become easy to solve.
No, I mean that there appears to be a basis of a solution already. Actually, multiple solutions being pursued by different groups. It's like asking a rocket engineer in 1955 how to build a rocketship to the Moon, or a physicist in 1936 how to build an atomic bomb, or the Wright brothers in 1900 how to build an airplane. Sure, in every single one of these cases you wouldn't get an exact, definitive answer. The Wright brothers didn't even understand the aerodynamics of their airplane, for example. But there were known avenues of inquiry for which there was very solid reason to believe that they would not be dead ends.
We're at a point now with AI/ML where solutions can be learned by machines for any solvable problem. It just needs some humans doing the selection of algorithms and guiding in hyper parameter space. But there is active research on automating that meta level which is yielding results. And both the reinforcement learning and the older AGI communities have working, tested designs for cognitive architectures that are truly general.
I'm not claiming we're done. I'm just saying that we're basically at the level of a working Wright flyer -- a bunch of research projects individually exhibit intelligence in separate domains, and a couple of cognitive architectures for generalizing them which have been shown to work on toy problems. There's no known unknowns that would cause these approaches to fail, so the reasonable expectation is that in the coming decades we will see the rise of useful AGI. Just like a reasonable observer with all the facts in 1905 should have predicted consumer passenger air travel.
> But we don't have evidence of intelligence arising that way in the past. What we have is the human brain, an incredibly complex structure...
Yes the human brain is difficult to understand. So is the flight of a bird. It's a good thing that we don't need to replicate the mechanics of bird flight to build flying machines -- otherwise you and I would still be stuck to trains and boats for getting around.
I suggest looking not at a neuroscience text book but a psychology textbook. Ask yourself not whether you can replicate exactly the conditions going on in the brain, but rather ask if you can implement a program to the same general functional description as the psychology textbook provides. That's a much easier task, and one well within the capabilities we have today.
A good example is Google Translate. We stopped trying to "understand" language and just threw a large dataset at pattern matching. And it worked amazingly well.
For other things we throw ideas of "this might improve what we can already do" and it winds up effectively solved, whether or not we understand it. This happened, for example, with Alpha-Go.
So "conceptually solved" and "AI can do it" are pretty much disconnected.
The genius of the RL feedback loop makes me think we can get to AGI without understanding human intelligence holistically.
To me google translate seems like the perfect illustration of the shortcomings of this purely data driven approach...
Alpha go, on the other hand is a much much more elaborated algorithm ( from my very limited knowledge) and it does produce amazing results.
Yes, I can list plenty of cases where translation fails. For example not understanding that in a conversation about specialist topic X (eg Ruby programming, or the game of G) that seemingly common terms have specific meanings. But I can also use it on a random newspaper article in a different language and expect to understand the result. This was something that I did not expect to happen in my lifetime.
Suppose a team of reseachers were given a computing platform a trillion times more powerful than the best systems available today, and a trillion dollars to build a set of training data. Would they be able to create a general AI in a predictable timeframe? If not then there's at least some area where we need to develop conceptual understanding. My intuition says that this is the case.
If you think that general AI can, in fact, be developed the same way as google translate, then what do you think the training data would look like? I'd imagine that the translation training set looks something like a massive Rosetta stone, though I haven't looked into it. Inputs associated with a valid output. What would a set of input/output pairs look like for a general AI? Any valid English query/statement/remark/essay with an "intelligent" response? Like a chatterbot? Doesn't seem like that would work. Or maybe I'm wrong. Maybe a chatterbot with ungodly amounts of computational resources and training data would develop the ability to reason. Maybe I'm just not appreciating how much that could change the game.
But it seems more likely that the first general AI will look more like Alphago, where we started with a reasonably solid understanding of how to approach the game, and then introduced deep learning where we discovered it was the best approach. We don't have a reasonably solid understanding of how to organize a general intelligence, and we don't know what it would take to get there.
Except that we don't. The original Google Translate used the conceptual modeling approach to translation, and it was absolute shit for any non-closely related language pairs. Google Translate (and Baidu, etc.) now uses deep networks whose operations are rather opaque. We don't understand in detail how they work. But that didn't stop us from building them.
It's not a useless notion of "conceptual understanding" though, because while being able to have a program complete a task implies having the conceptual understanding necessary for that task, it's possible to have the conceptual understanding but not be able to make a program, if, for instance, it would require greater (within reason) computing resources.
In particular Google spent around $100 million in 2003 for Applied Semantics, in part hoping that their natural language processing expertise could be used for translation. That technology was superseded by statistical techniques.
Computing increases have enabled a ton more experimentation with AI. Experiments that would have taken years to do, can be done in hours now. Experimentation lets scientists develop intuition about their algorithms. And get a sense of what will and won't work. And have a better mental model of the problem.
Not to mention the benefit of having 10x or more funding and researchers working on it now that results have been shown.
It's a common myth that there has been little or no innovation in NNs and that it's just a matter of computing power. Taking the best algorithms of the 90s and running them on modern hardware would still give you poor results. For instance, most of the 90s research was on very shallow nets, they didn't know how to properly train deep ones.
What has improved is the variety of architectures, but I don't think there has been anything fundamentally new after LSTMs. The only exception I can think of is generative adversarial networks, and even ideas behind adversarial training have been around in other domains for many years now.
The hierarchy is crucial: as in the brain, there are different regions at various levels of cognition, with higher regions responsible for greater levels of abstraction. A moment of real learning/understanding is characterized by a sudden, system-wide switch of signals between layers mostly traveling up the hierarchy (this input is confusing! I don't know what to do with it!) to signals mostly cascading down (AHA! I get what that means! Let's do this in response).
It was written in 2004, so like this article it lacks some of the insight we've gained over the past decade or so. But I think some of its general insights are still relevant. Whether it's NNs or HTMs or some yet-to-be-discovered algorithm that ends up being the driver of general AI, this article definitely resonates. At the most general levels of AI, we are still in the "confused" phase: viewing the state of AI as a whole system, our signals are still largely questions going up the conceptual hierarchy, rather than understanding cascading down.
Of course, what's missing from hypothetical artificial arithmetic is simple symbolic manipulation. And I am pretty sure that "More gofi", a more explicit treatment of AI, is something that Noam Chomsky and other have as their favorite bullet point on how to go further.
However, that too has been tried to whatever degree. If you had an overt, tractable logical theory that explained all intelligent behavior, then yes you'd have the missing ingredient of human intelligence. But unlike arithmetic calculation, it seems unlikely that human intelligence has this quality.
Indeed, all human "dealing with the world" behavior together involves a black box whose entirety is not subject to rational or logical reflection or description but whose broad outline. A person can't give a complete-enough-to-write-an-algorthim account of walking down the street recognizing things but the person can likely give a good why the street light they see is a street light - ie, heuristic black box behavior and logical/deductive behavior is intimately tied within human behavior.
And this might give some clue what's missing modern AI.
Unlikely as Minsky (the prince of gofai) and Chomsky were, umm, hardly fans of each other's work.
What was the compact, correct way referring to here?
either this "label" has no place in the story or it is not meant to be part of the story, but then our number-words do have an inherent structure, so they are easy to understand.
The mathematically innumerate comments are almost a troll that if the thought experiment worked and an artificial mathematician were invented, a significant fraction of the real world would not recognize or agree with its results.
There's a very cruel saying about people should do what they're born to do, with the dark insinuation that grandma the knitter should be locked in a sweatshop to sew against her will or (insert trendy CS tech here) children should be euthanized as a general policy until the tech gains significant commercial traction. Anyway unless an AI is quite heavily socialized into our culture what we're likely to grow might be ideally suited to research Klingon Warp Engine Fields in the 24th century but to us the output is going to look like hard to compress digital noise, so there's that problem.