Daniel M. Berry

Cheriton School of Computer Science
University of Waterloo
Waterloo, ON, Canada

Abstract:

It is well known that correcting as little as 10% of a program written by people in one's own project can cost again the cost of writing the program manually, because the cost to correct any code is about 10 times the cost of writing it correctly from the beginning. Typically, half of a program generated by an LLM has to be corrected, far more than for a program generated by people. This correction can cost 5 times the cost of writing the program manually. In addition, before the people can begin to correct an LLM-generated program, they must study the program to learn what the LLM did. Therefore, I hypothesize that writing a correct program with AI assistance will be significantly more expensive than writing the same program manually. The talk shows reports, blogs, and posts reporting experiences consistent with the hypothesis.