DEREK ZAHNS Thu 10/18/2007 11:45 AM POST STRUCK ME AS RAISING SOME PARTICULARLY IMPORTANT QUESTIONS. MY RESPONSES ARE IN ALL-CAPS.
> 1. What is the single biggest technical gap between current AI and AGI? I think hardware is a limitation because it biases our thinking to focus on simplistic models of intelligence. However, even if we had more computational power at our disposal we do not yet know what to do with it, and so the biggest gap is conceptual rather than technical. I THINK SOME PEOPLE, SUCH AS THOSE AT NOVAMENTE, HAVE SOME PRETTY GOOD IDEAS OF IMPORTANT THINGS TO TRY. In particular, I become more and more skeptical that the effort to produce concise theories of things like knowledge representation are likely to succeed. Frames, is-a relations, logical inference on atomic tokens, and so on, are efforts to make intelligent behavior comprehensible in concisely describable ways, but they seem to only be crude approximations to the "reality" of intelligent behavior, which seem less and less likely to have formulations that are comfortably within our human ability to reason about effectively. As one example, consider the study in cognitive science of the theory of categories -- from the "necessary and sufficient conditions" classical view to the more modern competing views of "prototypes" vs "exemplars". All of these are nice simple descriptions but as so often happens it seems that the effort to boil down the phenomena to nice simple ideas we can work with in our tiny brains actually boils off most of the important stuff. The challenge is for us to come up with ways to think about or at least work with (and somehow reproduce or invent!) mechanisms that appear not to be reduceable to convenient theories. I expect that our ways of thinking about these things will evolve as the systems we build operate on more and more data. As Novamente's atom table grows from thousands to millions and eventually billions of rows; as cortex simulations become more and more detailed and studyable; as we start to grapple with semantic nets containing many millions of nodes -- our understanding of the dynamics of such systems will increase. Eventually we will become comfortable with and become more able to build systems whose desired behaviors cannot even be specified in a simple or rigorous way. THE ABOVE TWO PARAGRAPHS ARE VERY INTERESTING. THEY ARE WHAT MOTIVATED ME TO WRITE THIS RESPONSE. YOU ARE PROBABLY CORRECT THAT THE REPRESENTATIONS NEEDED FOR HUMAN LEVEL AGI WILL BE COMPLEX AND NOT EASILY REDUCED TO THINGS OUR BRAINS CAN EASILY DEAL WITH. MUCH OF THIS PROBLEM COMES FROM THE FACT THAT THE EXPERIENCE OF EXTERNAL REALITY THE BRAIN IS TRYING TO MODEL IS ITSELF VERY COMPLEX AND NOT EASILY REDUCIBLE INTO A SIMPLE FRAMEWORK. AS IN NOVAMENTE, COMPLEX THINGS WILL BE REPRESENTED BY COMPLEX NETS, AND WHICH ASPECTS OF THAT COMPLEXITY ARE OPERATIVELY RELEVANT AT A GIVEN TIME MUST BE ABLE TO CHANGE RAPIDLY IN A CONTEXT AND TASK APPROPRIATE MANNER, WHICH ADDS FURTHER COMPLEXITY. BUT IT IS NOT CLEAR TO ME THIS MEANS COMING UP WITH WAYS FOR DEALING WITH SUCH COMPLEXITY ARE CURRENTLY DECADES BEYOND US. IF THERE WERE SIGNIFICANT FUNDING FOR EXPERIMENTING WITH THE TYPES OF LARGE COMPLEX REPRESENTATIONS YOU TALK ABOUT ABOVE, UNDERSTANDING THE DYNAMICS OF SUCH SYSTEMS MIGHT COME MUCH MORE QUICKLY. REMEMBER THE GOAL IS TO HAVE THE SYSTEM LEARN MOST OF THE COMPLEXITY NEEDED, PARTICULARLY IN TERMS OF REPRESENTATION AND BEHAVIORS, BUT ALSO IN TERMS OF PARAMETER SETTING AND, TO SOME EXTENT, IN TERMS OF ALGORITHMS. SO THE COMPLEXITY WE HAVE TO MASTER TO BUILD SUCH A HUMAN-LEVEL INTELLIGENCE IS MUCH LESS THAN THE COMPLEXITY OF A HUMAN MIND, OR THE COMPLEXITY OF THE REPRESENTATION AND OPERATION OF THE MACHINE, ITSELF. Or, perhaps, theoretical breakthroughs will occur making it possible to describe intelligence and its associated phenomena in simple scientific language. I DOUBT IT (AS I PRESUME YOU DO). THE BASIC ARCHITECTURE OF THE SYSTEM, IN TERMS OF WHAT IS PRE-PROGRAMMED INTO IT MIGHT BE RELATIVELY SIMPLE, AND MANY OF ITS BASIC OPERATIONS MIGHT BE RELATIVELY SIMPLE, AND IT MAY BE ABLE TO COMMUNICATE TO US THE THINGS IT IS THINKING AS WELL AS A HUMAN, AND WE MIGHT DEVELOP SOME VERY USEFUL SIMPLIFICATIONS AND GENERALIZATIONS, BUT WORLD KNOWLEDGE IS GOING TO BE A TANGLED MESS OF INTERCONNECTIONS -- MILLIONS OR BILLIONS OF THEM -- SOME OF WHICH MAY HAVE LITTLE CLEARLY DEFINABLE MEANING TO US HUMANS. THE DYNAMIC STATE INSIDE A HUMAN-LEVEL AGI, OVER EVEN A FEW SECONDS, WILL BE EVEN MORE COMPLEX. Because neither of these things can be done at present, we can barely even talk to each other about things like goals, semantics, grounding, intelligence, and so forth... the process of taking these unknown and perhaps inherently complex things and compressing them into simple language symbols throws out too much information to even effectively communicate what little we do understand. I THINK WE CAN CURRENTLY MEANINGFULLY TALK TO EACH OTHER ABOUT THINGS LIKE GOALS, SEMANTICS, GROUNDING, INTELLIGENCE... YES, OUR UNDERSTANDING OF THEM WILL BE MUCH BETTER WITH THE TYPE OF UNDERSTANDING YOU SO PROPERLY ADVOCATE ABOVE, UNDERSTANDING YOU SUGGEST WE MAY GET FROM EXPERIENCE WITH LARGE SYSTEMS. BUT WE CAN SAY MEANINGFUL THINGS ABOUT THESE CONCEPTS NOW, AND SUCH MEANINGFUL THOUGHTS ARE IMPORTANT FOR GENERATING THE LARGE SYSTEMS WE NEED TO GET THE UNDERSTANDING YOU TALK ABOUT. EXPERIMENTING WITH SUCH CONCEPTS IN SUCH LARGE SYSTEMS SHOULD BE AN IMPORTANT PART OF THAT LEARNING EXPERIENCE Either way, it will take decades if we're lucky. Moving from mouse-level hardware to monkey-level hardware in the next couple decades will be helpful, just like our views on machine intelligence have expanded beyond those of our forebears looking at the first digital computers and wondering about how they might be made to think. IT MAY WELL TAKE A DECADE, BUT I DOUBT IT WILL TAKE TWO. THAT IS, IF THERE IS ANY FUNDING THAT IS AT ALL COMMENSURATE WITH THE EXTREME IMPORTANCE OF THE FIELD. IT IS ALMOST CERTAIN THAT HUMAN-BRAIN-LEVEL HARDWARE COULD BE PROFITABLY SOLD FOR BETWEEN SEVERAL HUNDRED THOUSAND AND SEVER MILLION DOLLARS IN 10 YEARS, AND IT IS HIGHLY LIKELY THAT SYSTEMS 1/10 THAT SIZE WOULD BE VERY USEFUL INTELLIGENCES FOR MANY TASKS AND A GOOD TEST BED FOR THE TYPE OF LEARNING (AND PRESUMABLY EXPERIMENTING) YOU SEEM TO BE SUGGESTING. YOU ARE VERY RIGHT IN THINKING OUR VIEWS ON MACHINE INTELLIGENCE HAVE MOVED WAY BEYOND THOSE FIRST THOUGHTS ABOUT HOW TO MAKE COMPUTERS THINK. YOU ARE ALSO CORRECT IN THINKING OUR UNDERSTANDING STILL HAS A WAY TO GO. Edward W. Porter Porter & Associates 24 String Bridge S12 Exeter, NH 03833 (617) 494-1722 Fax (617) 494-1822 [EMAIL PROTECTED] -----Original Message----- From: Derek Zahn [mailto:[EMAIL PROTECTED] Sent: Thursday, October 18, 2007 11:45 AM To: agi@v2.listbox.com Subject: RE: [agi] Poll > 1. What is the single biggest technical gap between current AI and AGI? I think hardware is a limitation because it biases our thinking to focus on simplistic models of intelligence. However, even if we had more computational power at our disposal we do not yet know what to do with it, and so the biggest gap is conceptual rather than technical. In particular, I become more and more skeptical that the effort to produce concise theories of things like knowledge representation are likely to succeed. Frames, is-a relations, logical inference on atomic tokens, and so on, are efforts to make intelligent behavior comprehensible in concisely describable ways, but they seem to only be crude approximations to the "reality" of intelligent behavior, which seem less and less likely to have formulations that are comfortably within our human ability to reason about effectively. As one example, consider the study in cognitive science of the theory of categories -- from the "necessary and sufficient conditions" classical view to the more modern competing views of "prototypes" vs "exemplars". All of these are nice simple descriptions but as so often happens it seems that the effort to boil down the phenomena to nice simple ideas we can work with in our tiny brains actually boils off most of the important stuff. The challenge is for us to come up with ways to think about or at least work with (and somehow reproduce or invent!) mechanisms that appear not to be reduceable to convenient theories. I expect that our ways of thinking about these things will evolve as the systems we build operate on more and more data. As Novamente's atom table grows from thousands to millions and eventually billions of rows; as cortex simulations become more and more detailed and studyable; as we start to grapple with semantic nets containing many millions of nodes -- our understanding of the dynamics of such systems will increase. Eventually we will become comfortable with and become more able to build systems whose desired behaviors cannot even be specified in a simple or rigorous way. Or, perhaps, theoretical breakthroughs will occur making it possible to describe intelligence and its associated phenomena in simple scientific language. Because neither of these things can be done at present, we can barely even talk to each other about things like goals, semantics, grounding, intelligence, and so forth... the process of taking these unknown and perhaps inherently complex things and compressing them into simple language symbols throws out too much information to even effectively communicate what little we do understand. Either way, it will take decades if we're lucky. Moving from mouse-level hardware to monkey-level hardware in the next couple decades will be helpful, just like our views on machine intelligence have expanded beyond those of our forebears looking at the first digital computers and wondering about how they might be made to think. _____ This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/? <http://v2.listbox.com/member/?& > & ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=55050483-f80d2c