Superintelligence 2017-2056

Superintelligence is transforming the way people live and interact with the world around them. Our SIs are getting better, and things are improving faster than we think possible.

The amount of work our SIs can finish exceeds what the unaided human alone can do, literally compressing months of effort into a single night of SI-assisted work. That capacity doubles about every four and a half months, and it is increasing exponentially. SIs have started writing their own code, running their own training, and designing their successors. About a year and a half of progress is already being made each calendar year, and it continues to accelerate. In thirty short years, it will be like we all have superpowers of our own.

Before superintelligence, there were other supers. Superman, the supersonic jet, the supercomputer: each one amplified a single thing. In 2017 a research paper introduced the transformer, and with it something new, a machine that amplifies intelligence itself. By 2026 it had begun to speed up its own development.

Superintelligence 2017-2056 measures that acceleration with public data and explains it in plain language. It answers four questions.

Where did it begin? With the transformer, which this book argues was the most consequential discovery in human history, because it amplifies intelligence itself, the source of every other discovery.

How fast is it moving, and why is it speeding up? Faster than at any time in its history, measured in a reanalysis of sixty-four thousand evaluation runs across seven laboratories, and faster every year because superintelligence now does a growing share of the work that produces the next superintelligence.

Why can most people not see it? Because they are busy, because human intuition extends change in a straight line, and because compounding change looks like a straight line for years before it bends.

How does superintelligence come to build itself? Loop by loop. It learns to catch its own repeated failures by about 2029, builds software from a specification by 2036 and whole business systems by 2046, and by 2056, the year of the Singularity, it designs, trains, and deploys its own successors.

Along the way the book names the eleven failures that still keep a person in the loop and shows how each one is being fixed years early.

This is the third book in a series. The Singularity Roadmap 2026-2116 is the map of the road, and The Singularity of 2056 is the portrait of the destination. This book explains the engine: why the vehicle keeps going faster.

With 166 figures, and a dated test for every forecast.