¶ Rudolph's Cud · 22 July 2006 essay/tech
I spent this week at the annual conference of the American Association for Artificial Intelligence. I have no academic background in artificial intelligence. Of course, I have no academic background in computer science, or interface design, or technical support, or pretty much anything else I've ever been paid to do. The tenuous premise of my professional career is that I think well, in general, and maybe that I think particularly well in the intersection between the things that might improve human lives and the things that machines can do. My track record indicates that I've been involved with projects where this tension has been usefully resolved, and my day-to-day work seems to support my presumptuous contention that I am a personally significant contributor towards such resolutions.
So I'm clearly unqualified to assess, in general, the internal quality of advanced research work in the field of artificial intelligence. But I'm in charge of a software project in which machine-learning algorithms are being put to practical use, so I have a stake in the field. And the general meta-question of the application of technology to the class of what, to humans, are thinking problems, is probably no farther outside of my domain than anything else I ever deal with. And even if I were deeper in the field, there were six parallel tracks of talks taking place during most of the conference, so no one person could really hope to directly evaluate the whole thing anyway.
Take this, then, for whatever it evokes. My scattered impressions cluster around two ideas. The first, which covers about half of my experiences during the week, is that I had wandered into a convocation of alchemists. Time after time I sat through earnest descriptions of patently clever mechanisms for conveying the lead to the transmutation chamber, or of organizing the variations of gold sure to soon be produced, or of venting the toxic fumes that somebody had proved (in last year's paper) would result from the conversion. Not so much actual gold, just yet. This year's conference was celebrating the 50th anniversary of AI as a field, and although it would be wrong to dismiss a hard problem on the grounds that it hasn't been solved quickly, the only way Zeno got away with diminishing returns was by publishing measurable progress so frequently.
One major branch of AI is based on the observations that humans think, and that we express thoughts, and that our expression of each seemingly individual thought involves a complex background of assumptions and definitions and relationships and conceptual leaps, so maybe if computers had all that context to work in, they'd be able to think, too. The logistical and emotional center of the cargo-cult effort to build that foundation of knowledge, and hope it attracts Thinkingness, is the Cyc project, whose director proudly told a room full of people that they have so far encoded 1,000,000 terms, 100,000 relationships between them, and 10,000 meta-assertions about those things and relationships, and that they believe this constitutes about 10% of what is required.
Then he added "(isa Rudolph reindeer)" as 1,000,001 and 10,001, and with a flourish derived "Rudolph is a ruminant", which is so far from 10% correct that I'm still laughing about it a week later. I suspect the Atlantean Association of Alchemical Investigators thought they were 10% of the way to transmutation at some point, too. To me this is not only almost certainly wrong, but also fabulously self-ridiculing. There is no reasonable way to assess the size of something you can't even define, and you can't be 10% to nowhere any more than you can be 10% to a miracle. 10% to a miracle is not a scientific progress-report, it's a numerology of the angel-capacity of certain ornamental pins.
And even if it made any sense to say that we're 10% of the way to encoding the basics of human common sense, to me the premise of the project is fatally flawed in three ways:
1. Our ability to carry on conversations is less a product of the depths of information we plumb with every statement than of our ability to levitate on ignorance. You can always add a new level of notation to explain what the previous level means, but you can always add a new level, so all you've really succeeded in doing is turning the second M in MLM to Modeling and wasting the time of a lot of people who already knew what you meant in the first place. The machines might remember everything you can figure out how to tell them, but remembering isn't half as useful as forgetting. Not only is deduction less valuable than guessing, it's ultimately even less valuable than being productively wrong. And if after 50 years of AI and 20 years of Cyc we're only 10% of the way to encoding the things people know, then you and I will be long dead before we ever so much as get started on the exponentially larger domain of our idiocies and misconceptions and errors.
2. Even if we could encode all of that, I'm in no way convinced that it's necessary for thinking, let alone sufficient. I sat through a lot of talks about what felt to me, ultimately, like attempts to figure out how to build a bicycle by taking apart a horse. Cyc is pretty much the embodiment of Intelligent Design as an engineering methodology, trying to build a grown human adult an atom at a time. I suspect they would defend themselves by saying that they're only really trying to build a 14-year-old, and then they can have it read the rest of the stuff in books. This is never going to work. Even building a baby and letting it learn the whole thing from scratch isn't going to work. Complexity evolves. If we're going to build artificial constructs that think, we're going to do it by sowing the most primitive of artificial creatures into the most accelerated of generative artificial worlds and giving them virtual epochs in which to solve (and state) their problems themselves.
3. Underlying both of these errors, I think, is one fundamental misconception, engendered by the words we use to describe the point of all this speculation and work, and maybe most encouraged to fester by the inescapable apparent simplicity of one unfortunate thought-experiment. It's easy to imagine talking to a computer. The genius of Turing's Test is that it requires less equipment than an E-meter audit and less training than teaching SAT-prep classes. It's even easier to understand in the IM era than it was when he posed it. If a computer can convince us it's a person, then it is thinking. Translated to what I suspect most people would think this means, and then inverted for logic, this basically says that "people" are whatever we can't tell aren't like ourselves. But this is Fake People, not Artificial Intelligence, and shouldn't be what we mean by "thinking", or by artificial intelligence. My cats apply the Turing Test to me several times a day. I always fail. The first machine to think will be the first one that comes up with its own idea of what thinking means. It won't be like us. It won't think like us, it may not have much to say to us, and it almost certainly won't care. It will be good at things it likes to do, not the stuff we want just because we want it.
And therein, I think, is something closer to what I think we ought to mean by artificial intelligence: not the manipulation of ideas, however facile or reified, but the generation of them. Here's the Furia Test: the first true AI is the first machine to learn to spit out Rudolph's cud and tell the Turing judges to take their incessant idiotic questions and fuck off.
But I said that that was only half of my experience. Humans have come up with a lot of good ideas while working on bad ones, and arguably most human progress takes place in the context of grandly evocative error. I'm not really interested in building a machine I can condescend to, and I bet a lot of the other people at the conference weren't there as part of that dream, either. Taking apart a horse teaches you things. Building bicycles teaches you things. Even the effort to teach Cyc about ThingsThatMakeFartingNoisesWhenYouAccidentallySitOnThemFn teaches you something.
My part of this comes from a much, much simpler dream. I already know people who think. I don't need machines for that. True AI will have its own questions to devise and then answer. What I need is deeper and broader answers to human questions we already know, and new questions that matter to us. I see that the information we already have, jammed into computers that don't think and aren't likely to start thinking any time soon, is sufficient to answer several orders of magnitude more questions than our current question-answering tools facilitate. This isn't an existential problem, and it's our problem because it's so clearly our fault. Computers would be perfectly happy to synthesize and connect, but instead we've obtusely put them to work obscuring information from each other so that even the syntheses inherent in the character of the information become intractable. The machine-learning parts of the system I'm working on are tangibly useful, but in a sense mostly regrettable necessities, devised to undo the structural damage done by too-exclusively packaging small amounts of knowledge for unhelpful individual resale.
The good news from AAAI, for me, is that there are incredibly interesting things going on in the subfields of knowledge representation and reasoning. I actually do have some academic background in deductive logic, and the project I'm working on is both deeply and shallowly involved with information structure, so maybe my grounding in this subarea is just a little firmer. More than that, though, I can understand why it matters. Some of the new information-alchemy equipment is really new good information-chemistry gear opportunistically relabeled. I understand, for example, that we need graph-structures in most of the places where we have heretofore had only trees, and that our tools for comprehending graph structures are way behind our tools for understanding tree structures, and our tools for understanding tree structures without chopping them down first were already none too good. I understand what each order of logic adds, if that's not all you're doing, and what the effort to encode each new level costs and potentially loses. I understand that we want to make closed-world decisions in an open world, and that we have to. I understand why Tim Berners-Lee wants URIs on everything and I understand that a huge part of how we're able to talk to each other about anything is that we are not required to unambiguously identify every element of what we are attempting to say. I understand that Zeno was only ever half right, and that he never had to actually ship anything.
So I go back to work, where we aren't trying to build machines that think, and aren't waiting for anybody else to. Our bicycles aren't going to look or smell much like horses, but we're going to try our best to make them faster than walking. I am bemused to have fallen into the company of alchemists, but even at their most insularly foolish they're a lot more interesting to talk to than Sarbanes-Oxley compliance officers. I'd much rather worry about what OWL leaves out than have to explain to business-unit managers why writing "XML BPM" on a PowerPoint slide doesn't constitute an advanced-technology plan. Ultimately my complaint about AI is probably not so much that its premises are flawed as that the current particular flaws lead to wrong work that is too mundane in its wrongness. I understand that I am paid to be an applied philosopher, and I want to feel, after a week among people freer to linger in theory, that practice is pedaling flat-out just to keep possibility in sight. But if we're just out here with our new fancy bicycle prototypes, and our old blank maps, and no plausible rumors of buried gold or miracles, then fine. There are places to ride, and better things to see by looking up than down, and we'll see more of them if we aren't always stopping to dig new collapsing holes in the same empty sand.
So I'm clearly unqualified to assess, in general, the internal quality of advanced research work in the field of artificial intelligence. But I'm in charge of a software project in which machine-learning algorithms are being put to practical use, so I have a stake in the field. And the general meta-question of the application of technology to the class of what, to humans, are thinking problems, is probably no farther outside of my domain than anything else I ever deal with. And even if I were deeper in the field, there were six parallel tracks of talks taking place during most of the conference, so no one person could really hope to directly evaluate the whole thing anyway.
Take this, then, for whatever it evokes. My scattered impressions cluster around two ideas. The first, which covers about half of my experiences during the week, is that I had wandered into a convocation of alchemists. Time after time I sat through earnest descriptions of patently clever mechanisms for conveying the lead to the transmutation chamber, or of organizing the variations of gold sure to soon be produced, or of venting the toxic fumes that somebody had proved (in last year's paper) would result from the conversion. Not so much actual gold, just yet. This year's conference was celebrating the 50th anniversary of AI as a field, and although it would be wrong to dismiss a hard problem on the grounds that it hasn't been solved quickly, the only way Zeno got away with diminishing returns was by publishing measurable progress so frequently.
One major branch of AI is based on the observations that humans think, and that we express thoughts, and that our expression of each seemingly individual thought involves a complex background of assumptions and definitions and relationships and conceptual leaps, so maybe if computers had all that context to work in, they'd be able to think, too. The logistical and emotional center of the cargo-cult effort to build that foundation of knowledge, and hope it attracts Thinkingness, is the Cyc project, whose director proudly told a room full of people that they have so far encoded 1,000,000 terms, 100,000 relationships between them, and 10,000 meta-assertions about those things and relationships, and that they believe this constitutes about 10% of what is required.
Then he added "(isa Rudolph reindeer)" as 1,000,001 and 10,001, and with a flourish derived "Rudolph is a ruminant", which is so far from 10% correct that I'm still laughing about it a week later. I suspect the Atlantean Association of Alchemical Investigators thought they were 10% of the way to transmutation at some point, too. To me this is not only almost certainly wrong, but also fabulously self-ridiculing. There is no reasonable way to assess the size of something you can't even define, and you can't be 10% to nowhere any more than you can be 10% to a miracle. 10% to a miracle is not a scientific progress-report, it's a numerology of the angel-capacity of certain ornamental pins.
And even if it made any sense to say that we're 10% of the way to encoding the basics of human common sense, to me the premise of the project is fatally flawed in three ways:
1. Our ability to carry on conversations is less a product of the depths of information we plumb with every statement than of our ability to levitate on ignorance. You can always add a new level of notation to explain what the previous level means, but you can always add a new level, so all you've really succeeded in doing is turning the second M in MLM to Modeling and wasting the time of a lot of people who already knew what you meant in the first place. The machines might remember everything you can figure out how to tell them, but remembering isn't half as useful as forgetting. Not only is deduction less valuable than guessing, it's ultimately even less valuable than being productively wrong. And if after 50 years of AI and 20 years of Cyc we're only 10% of the way to encoding the things people know, then you and I will be long dead before we ever so much as get started on the exponentially larger domain of our idiocies and misconceptions and errors.
2. Even if we could encode all of that, I'm in no way convinced that it's necessary for thinking, let alone sufficient. I sat through a lot of talks about what felt to me, ultimately, like attempts to figure out how to build a bicycle by taking apart a horse. Cyc is pretty much the embodiment of Intelligent Design as an engineering methodology, trying to build a grown human adult an atom at a time. I suspect they would defend themselves by saying that they're only really trying to build a 14-year-old, and then they can have it read the rest of the stuff in books. This is never going to work. Even building a baby and letting it learn the whole thing from scratch isn't going to work. Complexity evolves. If we're going to build artificial constructs that think, we're going to do it by sowing the most primitive of artificial creatures into the most accelerated of generative artificial worlds and giving them virtual epochs in which to solve (and state) their problems themselves.
3. Underlying both of these errors, I think, is one fundamental misconception, engendered by the words we use to describe the point of all this speculation and work, and maybe most encouraged to fester by the inescapable apparent simplicity of one unfortunate thought-experiment. It's easy to imagine talking to a computer. The genius of Turing's Test is that it requires less equipment than an E-meter audit and less training than teaching SAT-prep classes. It's even easier to understand in the IM era than it was when he posed it. If a computer can convince us it's a person, then it is thinking. Translated to what I suspect most people would think this means, and then inverted for logic, this basically says that "people" are whatever we can't tell aren't like ourselves. But this is Fake People, not Artificial Intelligence, and shouldn't be what we mean by "thinking", or by artificial intelligence. My cats apply the Turing Test to me several times a day. I always fail. The first machine to think will be the first one that comes up with its own idea of what thinking means. It won't be like us. It won't think like us, it may not have much to say to us, and it almost certainly won't care. It will be good at things it likes to do, not the stuff we want just because we want it.
And therein, I think, is something closer to what I think we ought to mean by artificial intelligence: not the manipulation of ideas, however facile or reified, but the generation of them. Here's the Furia Test: the first true AI is the first machine to learn to spit out Rudolph's cud and tell the Turing judges to take their incessant idiotic questions and fuck off.
But I said that that was only half of my experience. Humans have come up with a lot of good ideas while working on bad ones, and arguably most human progress takes place in the context of grandly evocative error. I'm not really interested in building a machine I can condescend to, and I bet a lot of the other people at the conference weren't there as part of that dream, either. Taking apart a horse teaches you things. Building bicycles teaches you things. Even the effort to teach Cyc about ThingsThatMakeFartingNoisesWhenYouAccidentallySitOnThemFn teaches you something.
My part of this comes from a much, much simpler dream. I already know people who think. I don't need machines for that. True AI will have its own questions to devise and then answer. What I need is deeper and broader answers to human questions we already know, and new questions that matter to us. I see that the information we already have, jammed into computers that don't think and aren't likely to start thinking any time soon, is sufficient to answer several orders of magnitude more questions than our current question-answering tools facilitate. This isn't an existential problem, and it's our problem because it's so clearly our fault. Computers would be perfectly happy to synthesize and connect, but instead we've obtusely put them to work obscuring information from each other so that even the syntheses inherent in the character of the information become intractable. The machine-learning parts of the system I'm working on are tangibly useful, but in a sense mostly regrettable necessities, devised to undo the structural damage done by too-exclusively packaging small amounts of knowledge for unhelpful individual resale.
The good news from AAAI, for me, is that there are incredibly interesting things going on in the subfields of knowledge representation and reasoning. I actually do have some academic background in deductive logic, and the project I'm working on is both deeply and shallowly involved with information structure, so maybe my grounding in this subarea is just a little firmer. More than that, though, I can understand why it matters. Some of the new information-alchemy equipment is really new good information-chemistry gear opportunistically relabeled. I understand, for example, that we need graph-structures in most of the places where we have heretofore had only trees, and that our tools for comprehending graph structures are way behind our tools for understanding tree structures, and our tools for understanding tree structures without chopping them down first were already none too good. I understand what each order of logic adds, if that's not all you're doing, and what the effort to encode each new level costs and potentially loses. I understand that we want to make closed-world decisions in an open world, and that we have to. I understand why Tim Berners-Lee wants URIs on everything and I understand that a huge part of how we're able to talk to each other about anything is that we are not required to unambiguously identify every element of what we are attempting to say. I understand that Zeno was only ever half right, and that he never had to actually ship anything.
So I go back to work, where we aren't trying to build machines that think, and aren't waiting for anybody else to. Our bicycles aren't going to look or smell much like horses, but we're going to try our best to make them faster than walking. I am bemused to have fallen into the company of alchemists, but even at their most insularly foolish they're a lot more interesting to talk to than Sarbanes-Oxley compliance officers. I'd much rather worry about what OWL leaves out than have to explain to business-unit managers why writing "XML BPM" on a PowerPoint slide doesn't constitute an advanced-technology plan. Ultimately my complaint about AI is probably not so much that its premises are flawed as that the current particular flaws lead to wrong work that is too mundane in its wrongness. I understand that I am paid to be an applied philosopher, and I want to feel, after a week among people freer to linger in theory, that practice is pedaling flat-out just to keep possibility in sight. But if we're just out here with our new fancy bicycle prototypes, and our old blank maps, and no plausible rumors of buried gold or miracles, then fine. There are places to ride, and better things to see by looking up than down, and we'll see more of them if we aren't always stopping to dig new collapsing holes in the same empty sand.