Professor Sir Tony Hoare has proposed an alternative Turing Test that could yield more effective tools for software developers at a prestige conference dedicated to the developer of arguably the famous method for assessing the quality of artificial intelligence.
Speaking at the Alan Turing Centenary Conference in Manchester, Hoare, principal researcher at Microsoft Research in Cambridge, asked whether computers can understand their own programs – and how long it might take for that to become possible.
“Turing predicted that within 50 years that a machine would pass his test: it seems a little nearer now than when he made that prediction," Hoard said. "If I had to make a 50-year prediction, this would be it: that a computer could understand its own program and assist the programmer to make changes and improvements,”
Hoare added: “This engineering version of the Turing Test would involve an interactive examination of the machine on a single topic – its own program – for as long as desired. The machine would not have to answer all of the questions. For his test, Turing showed that the computer will always have some questions it can’t answer.”
Hoare said the computer would be expected to answer questions such as whether the program it is running could overflow a buffer. “And if so, give a test case that shows the error. It would also be able to generate test cases that exercise all the changes recently made to the program and answer whether a certain change would make the program slower.”
Hoard predicted that "programs of the future will be written by collaboration, taking the particular skills of each of the parties [human and computer]. The human understands the real world, who will use the program and its commercial value. The computer understands the detail of the program and the consequences of changing it.”
Hoare explained that progress is being made to towards this end as computers are increasingly being called upon not just to demonstrate mathematical proofs but generate them. An example is the attempt to prove the Kepler conjecture on the most efficient packing of spheres. “The FlySpeck project aims to provide a completely computerised proof. In this work, there is a collaboration between human and computer, but it is very asymmetric,” he said.
Hoare stressed that understanding does not mean the computer would be able to reflect on the meaning of the program in the way a human would. “My version of the Turing Test doesn’t give much support to Alan Turing’s hope that we could make a computer that thinks.”