The Singularity Is Nearer: When We Merge with AI

by Ray Kurzweil

The most fundamental trend at work here is the exponentially improving price-performance of computation—that is, how many computations per second can be performed for one inflation-adjusted dollar. When Konrad Zuse built the first working programmable computer, the Z2, in 1939, it could perform around 0.0000065 computations per second per 2023 dollar.[201] In 1965, the PDP-8 managed around 1.8 computations per second per dollar. When my book The Age of Intelligent Machines was published in 1990, the MT 486DX could achieve about 1,700. When The Age of Spiritual Machines appeared nine years later, Pentium III CPUs were up to 800,000. And when The Singularity Is Near debuted in 2005, some Pentium 4s were at 12 million. As this book goes to print in early 2024, Google Cloud TPU v5e chips are likely around 130 billion operations per second per dollar! And because this awesome (quintillions of operations per second for a large pod) cloud computing power can be rented for a few thousand dollars an hour by anyone with an internet connection—as an alternative to building and maintaining a whole supercomputer from scratch—the effective price-performance available to retail users with small projects is several orders of magnitude higher than that. Because cheap computation directly facilitates innovation, this macro trend has steadily continued—and it does not rely on any particular technological paradigm like shrinking transistors or increasing clock speeds.