"To them, language could be modeled like a game of chance. At any point in a sentence, there exists a certain probability of what might come next, which can be estimated based on past, common usage. ... each step along the way is random, yet dependent on the previous step—a hidden Markov model. A speech-recognition system's job was to take a set of observed sounds, crunch the probabilities, and make the best possible guess about the 'hidden' sequences of words that could have generated these sounds. To do that, the IBM researchers employed the Baum-Welch algorithm—codeveloped by Jim Simons's early trading partner Lenny Baum—to zero in on the various language probabilities. Rather than manually programming in static knowledge about how the language worked, they created a program that learned from data."