What are some tasks where AI exceeded human expectations?

There have been many tasks that people in the past thought would be very hard or impossible for AI, which AI can now do.

Chess

In an essay about The Turk (a fraudulent automaton which was purported to be able to play chess), Edgar Allan Poe posited that no pure machine could adapt to the unpredictable complexities of chess. While he was correct that The Turk was not a pure machine, his thesis that no machine could ever play chess was eventually proven wrong.

Go

When chess champions were defeated by Deep Blue in 1997, it was suggested that the game of Go would resist computers for a very long time1 due to its much larger number of possible positions. 20 years later, Deepmind’s AlphaGo (and later the more general AlphaZero) surpassed humans. It’s worth noting that even superhuman AI players are still vulnerable to adversarial attacks.

Solving CAPTCHAs

CAPTCHAs are visual tests that are meant to be hard for computers but easy for humans to complete. These tests have repeatedly needed to be redesigned as computer algorithms improved, and LLMs seem to be able to solve many of them.

Turing tests

For some time, the Turing Test (which tests whether a human observer can distinguish a machine from a human based on textual outputs from each) was considered one major test of whether a machine could truly "think." It has been argued that LLMs have quietly passed the Turing Test and that we might need new forms of this test. (It’s worth noting that at least some humans have been fooled by computers pretending to be humans for over 50 years.)

Further reading:


  1. Shortly after Deep Blue beat Kasparov, Piet Hut said “It may be a hundred years before a computer beats humans at Go -- maybe even longer.” ↩︎