If you are at all concerned about a robot apocalypse and enslavement of the human race, do not read this. Researchers at MIT’s Computer Science and Artificial Intelligence Lab have created an AI program that learns how to read, to improve its success. To test the program it was asked to do two things. The first was to install a program, with only the instructions on Microsoft’s website to guide it. Its success won the team best-paper award at the Association for Computational Linguistics' 2009 meeting. The second test was to play the game Freeciv, an open source reimplementation of Civilization II, by only reading the manual for Civilization II. Games are often used for testing AI because of their highly complex nature and an opponent’s reaction to a player’s move will change randomly, preventing the player from simply memorizing all the reactions. When the AI, with access to the manual, played the game (MIT AI versus game AI) it was victorious 78.8% of the time, with a standard error of ± 5.8. The game's own AI which, when set against itself (game AI versus game AI), won only 45.7% of the time, with a standard error of ± 7.0. Obviously the MIT AI was more victorious, but the smaller standard error implies a greater consistency in the outcomes of the 50 full length games played.
Not bad considering the AI started playing before it could read. The research team is currently adapting their meaning-inferring algorithms to robotic systems.