This is an abstract of Edward Feigenbaum’s paper “Some Challenges and Grand Challenges for Computational Intelligence,” which presents a “grand vision” for the future of artificial intelligence research. You can read Feigenbaum’s own summary of the paper here.
A number of challenges may guide our research in artificial intelligence better than the “multidimensional” Turing test. The “Feigenbaum Test” compares the behavior of a human and a computer player based on their performance as experts in their discipline. If one in three of the expert judges fails to choose reliably between the two, the computer may be recognized as an expert system. Related challenges regard a machine’s ability to learn from what it “reads.” One challenge is to see whether, starting with a novice-level, symbolically represented understanding of a topic, a computer can read further texts in the field and encode (understand) that information. Human input on less than 10% of the reading would allow a dramatic increase in the rate of symbolically encoding knowledge. A computer able to do this could keep up on recent literature and repeat the Feigenbaum Test every two years. A further goal would be to semantically encode the knowledge spread across the Internet.