commuter would be trying to mimic the logical structure of german but now it's all correlation. >> well, that's exactly right. in the past we tried to teach the computer how language worked and tried to enshrine all of these things in software code is really hard whether it's a code or language. the new approach is to rely on data and statistics and turn the problem into a math problem. it's probability. what is the likelihood a word in one language is the best alternative to a word in another language. to do that you need to pour in lots of data, lots of different translations. >> every document you can find in both english and german and the computer just keeps looking at those and correlating. >> that's right. so every document would go in, for every book scan translayings from the libraries, they put that in as well. in the past when we tried to do that with millions of pages, it worked well but not great. google avails itself of billions