TransVision Wednesday Morning Session II
25 July 2007 (updated 15 August 2007)
I’m in Chicago for TransVision 2007. Here are some notes and thoughts from Wednesday morning session II.
Marvin Minsky spoke about problems that we need to solve. The most serious are population and longevity. Others include epidemics, pollution, energy sources, income-gap poverty, education, terrorism, biological diversity, cultural diversity, and catastrophes. He thinks uploading is the best solution. It provides a more manageable form. Who will work when number of retirees equals number of workers? Robots will be needed to maintain infrastructure. 2001 is past. Where is HAL? Many expert systems today rival human experts, but only work in a small domain. Most AI researchers have tried to find universal techniques for solving all kinds of problems. This was physics envy. AI divided into specialties. Most people in the field make money by writing statistical learning machines. Humans think about what they’ve been thinking. When we recognize a problem type, we activate a way to think. Emotions are a way to think. Human resourcefulness is based on using different ways to think. If you build systems with multiple ways to work then, if one fails, another can be tried. When our minds enter machines, our descendents will involve in unknown ways. We’ll need to escape the planet and ensure the universe remains meaningful. Scientists should celebrate new Sagan-like advocates.
Philip Rosedale talked about his background that led up to the development of Second Life. He gave a demo of Second Life. There is an island with reproducing animals and plants, made using the second life scripting language. People have built 3d environments wrong. The key is to build using an atomic model, and allow people to build things from the bottom up. Second life uses a server to simulate 12 acres. There are 12000 servers. Residents of world are creating content at 10 percent per month, which is faster than any one person can consume. Users tend to be adults, age 32 average with more than 40 percent female. There 40000 people that are cash flow positive in second life. Money and relations both play roles. Building can be done real time. Objects can obey physical laws, and a scripting language can customize them. 40,000,000 dollars per month worth of exchange happening in second life now. Instruments can be programmed and played. Only 30 percent of second life residents are in united states. A translator gadget can allow communication between persons of different languages. Second life is creating relationships and jobs.
Ben Goertzel spoke about attaining artificial general intelligence in virtual worlds. Embodiment is linked to what is happening in virtual worlds like second life. He is most interested in producing general intelligence. They have autonomy and a practical understanding of self and others. It has to understand the problem, as opposed to solving provided problems. This is contrasted to narrow AI. These are useful and can make money today. These include expert systems, gaming, etc. Narrow AI is generally of minimal use as a stepping-stone toward general AI. AGI-08.org in march 2008 will be first conference on artificial general intelligence. Deep blue, darpa grand challenge and google are narrow AI. These can’t adapt to new games, new types of vehicles or complex questions. These are quite different from making real thinking machines. A human could learn these other things with an explanation and some practice. We can’t create agi through incrementally generalizing narrow ai. We need to go back to basics. We need to create artificial simpletons, like babies and pets. We need to teach simpletons well, and convince millions of others to help teach them by making it fun. Then the practical applications will come. Where should the agi baby live? In a robot? In a virtual world? Without a body? If we want to understand it then we should have something in common with it. He has pursued development of agi in simulated worlds. The agi starts without object permanence. Psychological games are helpful, but the unpredictibility of the real world seems more helpful. Dogs in second life are narrow expert systems that cannot learn new tricks. It would be interesting to sell virtual babies in second life, as an annual subscription. Companies are doing job interviews in second life, but none can use an agi interviewer yet. The novamente project is aimed at creating agi. Learning algorithms include moses from metacog.org, and probabilistic learning systems. Cognitive development goes through infantile, concrete operational, formal, reflexive, and full self-modification. Why might novamente succeed where others have not? Our hardware is better. We have virtual worlds like second life with millions of teachers. The design is important, too. Start with animals, go through babies, and on to adults.