On Sat, Nov 22, 2014 at 9:24 AM, Tim Tyler via AGI <[email protected]> wrote: >> Typical software productivity is 10 lines per day or 2000 lines per >> person per year. So this is about a 270 person-year effort. A typical >> budget at $250K per person-year (Ph.D. level salary plus overhead) >> would imply a cost of $67 million. Using my original estimate of 300M >> lines for AGI, OpenCog is 0.2% complete. > > If we add in some runtime unit tests involving the Wikipedia corpus, > we should be able to get there within a week or so. Maybe beyond > human level - unless you also think that Wikipedia is 90% junk.
No, Wikipedia isn't code. It's training data. I am talking just about just the complexity of a human body and brain at birth. I didn't include the cost of the hardware and training. If you want to build a human brain sized neural network, you will also need a few petaflops of computing power. If you want to run experiments to test your software and you don't want them to run as long as it takes a child to grow up, then you will need even more computing power to run them faster. You need a lot of edit-compile-test cycles to develop 300M lines. Wikipedia is a nice database for text based applications like Watson, but it is hardly the size you need for AGI. If you want to train your vision system, you need a decade's worth of high resolution video, for example. Maybe you think 300M lines is way too high. You could argue that human DNA is 98% "junk" or non-coding. But it is there for a reason. We have about the same number of genes (20K) as the roundworm Caenorhabditis elegans, but it has only 100M bases as opposed to our 3B. Most of its DNA encodes proteins or RNA. But there is a problem with densely packed DNA, just like there is with code optimized for size. It is very hard to modify without breaking something. Humans are more highly evolved than bacteria eating nematodes because mutations in humans are more likely to be beneficial and less likely to be fatal. We can switch genes on and off between coding and non-coding, but C. elegans can't. It is a lot like a big software project where some of the code is unused, left over from older versions, but might be useful in a future version. It is much faster to re-enable it than to delete it and rewrite it through the slow process of evolution. There is a cost to carrying extra code, however. It is fairly small in humans. You have about 10^13 cells each with 6 pg (6B diploid bases) of DNA. The total is 60 g, or about 0.1% of your body mass. Copying DNA has an energy cost similar to synthesizing protein. Your body is 25% protein. This means that 0.4% of your food intake and metabolism is used to copy this mostly "junk" DNA. C. Elegans has about 1000 cells at the end of its larval stage when it is 650 um long and 40 um in diameter, or about 1000 ng weight. Its DNA is 0.2 ng. If its genome were human sized, it would be 30 times as large and require 2.4% of its metabolism. So there is greater evolutionary pressure to remove the unused DNA. DNA size varies widely between species. The genomes of lungfish and wheat are huge. But all mammals have at least 2B bases, which suggests a minimum size for evolution to work. Maybe you still disagree that 300M lines of code are required. If you said 100M lines I might agree. If you said 1M lines, then I would have to ask why after 60 years we are still paying people $70 trillion per year to do work that machines aren't smart enough to do. A new car (not self driving) has 30M lines of code. So if AGI were simpler than a car or an operating system, it should have been solved by now. Here are the market caps of the major players in AI who have petaflops of computing power and can afford $30 billion to write 300M lines of code. Google $365B Facebook $205B. IBM $159B I think this is our best hope for AGI. -- -- Matt Mahoney, [email protected] ------------------------------------------- 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
