“How Long Before Superintelligence?” is a 1998 paper by Nick Bostrom, written years before his 2014 book “Superintelligence” made the topic mainstream. It is an early, explicit attempt to put a date on machine intelligence that exceeds the human kind, and much of the framing Bostrom is now known for appears here first.
Bostrom defines superintelligence as “an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills.” He deliberately excludes collective entities like well-run companies or efficient bureaucracies, insisting the term means a single unified intellect at that level.
The paper’s prediction rests on a hardware argument. Bostrom estimates that matching human-level processing requires somewhere between 10^14 and 10^17 operations per second, then uses Moore’s Law to project when affordable computers will reach that range - roughly the first quarter of the 21st century, and possibly its first years. He treats software as the harder bottleneck, suggesting the necessary neuroscientific understanding of how the brain learns might take on the order of fifteen years, and argues that once human-level AI exists, “there will soon be superintelligence” through recursive self-improvement and continued hardware gains. A 2008 postscript walked back the confidence, with Bostrom assigning “less than a 50% probability to superintelligence being developed by 2033.”
The paper is a primary document for the modern superintelligence debate, sitting between I.J. Good’s 1965 “intelligence explosion” idea and the dense, book-length treatment Bostrom would publish in 2014.