I really enjoyed this piece, not because of the specific result, but because of the style of reasoning it represents. How much advantage, under what kind of rules, can be overcome with what level of intelligence?
Sometimes the answer is none. “I play x” overwhelms any level of intelligence at tic tac toe.
In larger and more open games the advantage of intelligence increases, because you can do more by being better at exploring the space of possible moves.
“Real life” is plausibly the largest and most open game, where the advantage of intelligence is maximized.
So, exploring the kind of question the OP posits can give us a kind of lower bound on how much advantage humans would need to defeat an arbitrarily smart opponent. And extending it to larger contexts can refine that bound.
By the time we hit chess-complexity, against an opponent not trained for odds games, we’re already at around two bishops odds for an uncommon-but-not-extreme level of human skill.
I think a lot of the problems that arise in discussing AI safety are a (in the best cases much more well reasoned) form of “You think an AI could overcome X odds? No way!” “Yes way!”
I really enjoyed this piece, not because of the specific result, but because of the style of reasoning it represents. How much advantage, under what kind of rules, can be overcome with what level of intelligence?
Sometimes the answer is none. “I play x” overwhelms any level of intelligence at tic tac toe.
In larger and more open games the advantage of intelligence increases, because you can do more by being better at exploring the space of possible moves.
“Real life” is plausibly the largest and most open game, where the advantage of intelligence is maximized.
So, exploring the kind of question the OP posits can give us a kind of lower bound on how much advantage humans would need to defeat an arbitrarily smart opponent. And extending it to larger contexts can refine that bound.
By the time we hit chess-complexity, against an opponent not trained for odds games, we’re already at around two bishops odds for an uncommon-but-not-extreme level of human skill.
I think a lot of the problems that arise in discussing AI safety are a (in the best cases much more well reasoned) form of “You think an AI could overcome X odds? No way!” “Yes way!”