Stanford Colloquium: Self-Improving Artificial Intelligence, Stephen Omohundro – 66 min – Oct 31, 2007.
We are on the verge of a radical new paradigm for both computer software and hardware. “Self-improving systems” will have detailed models of their own designs and will improve themselves by learning from their own operation. They will continuously adapt themselves to the tasks they need to perform. Eventually they will be able to improve every aspect of themselves: their programs, programming languages, specification logics, instruction sets, and hardware architectures. In this talk we present fundamental principles that underlie the operation of this kind of system. We show that they will be governed by a fundamental microeconomic theory first developed by von Neumann in 1944. This leads to a universal “Resource Balance Principle” by which they will optimally allocate resources to their subsystems, modules, and subprograms. It also provides the rational basis by which they will select the timing and amount of effort to devote to tasks like program compilation and data compression. Self-improvement of hardware will push toward reversible computation and atomically precise physical structures. We conclude with a discussion of some of the broader social implications of this kind of system.