On Tue, Dec 25, 2012 at 10:18 PM, Ben Goertzel <b...@goertzel.org> wrote: > Maybe OpenCog will make the breakthrough, maybe > someone else... But I bet the breakthrough > will NOT come from lossless text compression research ;-p ...
You understand that I use compression to measure language model accuracy. I am not proposing to use compression as a application for AGI. For most people, zip is good enough. Also, I'm sorry to hear about all your difficulties with funding and cancelled projects. I realize it is hard work and you have limited time. Thank you for correcting some of my misunderstandings about OpenCog. One purpose of my benchmark is to encourage research toward a common goal without imposing any restrictions on the methods used to reach it. Everybody has different ideas how to solve AGI. It must be very difficult to get a team to agree on a common methodology without clashing egos. You might know that Marcus Hutter and I disagreed on seemingly minor details of the benchmark, which is why there is a separate but very similar benchmark with prize money. I wanted no limits on time or hardware, but when you offer prize money you can't do that. I still got interest without prize money because people are willing to go to great lengths just to prove that they are right. I think that there is some valuable information from my research. For one, adding CPU and memory always helps. We don't know how much will be enough, but so far it is much higher than anyone thought. Second, the best algorithms are very complex. Loosely, the process seems to be to make random changes to the code and to keep the changes that improve compression. You end up with an incomprehensible mess. I would prefer to understand how my own code works, but it doesn't work that way. It's like looking at your own DNA. I would hate to build an AGI with no understanding of the process, but that's exactly how we make children. I still believe that you need to make incremental progress toward hard goals. That is how human civilization was made out of goop in 3 billion years on a planet sized molecular computer. A human brain is more complex than a human brain can understand, so it has to be a collective effort. You can't do it all yourself. And if people can't work together, then the goal has to be accomplished through competition. Groups are smarter than their members because their members disagree. I know this stuff is hard. I have been incrementally improving compression algorithms for 12 years. I've done thousands of experiments. Most of them didn't work. It took years just to figure out the best way to combine the predictions of different models. From 2000 to 2005 I was using a probability and a confidence (like PLN) before I tried weighted averaging in the logistic domain and discarding the confidence values. It wasn't obvious that this worked better until after months more of experiments, but it is generally accepted now. It was also 3 years before my compressors did well enough that anyone cared. At that point, other people started making improvements to it. It only went to the top of the benchmarks with free help from a lot of other people who wanted to prove their ideas. Making OpenCog open source is the smartest thing you can do in this regard. The other thing that needs to happen is for it to do something useful or interesting. Right now, I understand that it is a bear to install without errors, and then once I do it is far from obvious what I could do with it. I realize the software is still being written so I shouldn't expect it to do much. But maybe you have some ideas. -- -- Matt Mahoney, mattmahone...@gmail.com ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
