Will AI be able to think faster than humans?

Whether current AI systems should be considered to "think faster" than humans depends on the meaning of “think”. Computers add and multiply numbers at over a billion times human speed. Problems like chess and Go, which once seemed to require human-like thought, have also been reduced to mere calculations. But human thought can still solve many problems that current AI can’t solve at all.

Future AI, however, may “think” with human-level generality without being limited to human time scales. Just as current systems are orders of magnitude faster at arithmetic, future systems could become orders of magnitude faster at everything.

We know this because digital computers can operate at much greater serial speed than the brain. In a digital computer, signals propagate near the speed of light, millions of times faster than in the human brain, where they’re limited to 100 m/s. The brain can perform hundreds of serial steps per second; digital computers can perform billions.1

This means a computer program emulating a human brain could do all the same things much faster. It could experience millennia of subjective time, and do millennia of research or other cognitive work, in a single day.

Advanced AI probably won’t come about as a sped-up human brain, or even resemble one. But the possibility of such programs, seen as a lower bound on the limits to AI power, implies we may see systems that strategize and make theoretical breakthroughs at millions of times human speed or more. If not aligned correctly, such systems could cause a catastrophe while we humans were almost frozen in time.

  1. See Chapter 3, footnote 3 of Bostrom’s Superintelligence for details. ↩︎