Consider RYY : your best probabilistic guess of your next move.
Assuming you know yourself perfectly (or good enough to predict your own moves reliably at least), it will turn out RYY is very similar to you (RYY is not deterministic if you are not, but then it is still an opponent as skilled as you for all chess-related purposes).
Then since you have shown you win against RYK, I can guess that RYY will reliably win against RYK, which I find very surprising.
Why is emulating a stronger player less efficient than emulating yourself ? (It sounds more surprising to me formulated like that than how you said it.)
The explanation I see is that you don’t know Kasparov well enough to emulate him correctly (what you already pointed out) whereas you know yourself very well.
Then, the question that comes to my mind is : how can you use this knowledge to improve your play ?
I have received the advice “what would X do in your stead” by a number of people in a number of circumstances, including here by rationalists.
How can it be useful ?
If it is helpful, then it means that your cognition algorithm can be optimized and you know a specific way to improve it.
If you find yourself frequently finding that wondering about what someone else would do is helpful, then there is additional computation that could be saved if you knew how you know how that someone else behaves (by doing it directly instead of imitating) .
So, it’s purely a matter of knowing yourself : the advice I had received was no better than “think about yourself”.
The other question I wonder about is how that applies to artificial intelligence.
I don’t know more much about it. Is the “know yourself” important in that case ? How does an AI see its own source code ?
I guess the first step toward making this question meaningful would be to design precisely how the emulator works. A naive approach (that doesn’t account for styla and psychology) would be to use statistics for every position, and be random if an unprecedented position occurs.
Then it becomes clear that there is no “default player” to emulate. There is no such thing as the emulator emulating itself because the emulator is not a player. Bummer.
Consider RYY : your best probabilistic guess of your next move. Assuming you know yourself perfectly (or good enough to predict your own moves reliably at least), it will turn out RYY is very similar to you (RYY is not deterministic if you are not, but then it is still an opponent as skilled as you for all chess-related purposes).
Then since you have shown you win against RYK, I can guess that RYY will reliably win against RYK, which I find very surprising.
Why is emulating a stronger player less efficient than emulating yourself ? (It sounds more surprising to me formulated like that than how you said it.)
The explanation I see is that you don’t know Kasparov well enough to emulate him correctly (what you already pointed out) whereas you know yourself very well. Then, the question that comes to my mind is : how can you use this knowledge to improve your play ?
I have received the advice “what would X do in your stead” by a number of people in a number of circumstances, including here by rationalists. How can it be useful ? If it is helpful, then it means that your cognition algorithm can be optimized and you know a specific way to improve it. If you find yourself frequently finding that wondering about what someone else would do is helpful, then there is additional computation that could be saved if you knew how you know how that someone else behaves (by doing it directly instead of imitating) . So, it’s purely a matter of knowing yourself : the advice I had received was no better than “think about yourself”.
The other question I wonder about is how that applies to artificial intelligence. I don’t know more much about it. Is the “know yourself” important in that case ? How does an AI see its own source code ? I guess the first step toward making this question meaningful would be to design precisely how the emulator works. A naive approach (that doesn’t account for styla and psychology) would be to use statistics for every position, and be random if an unprecedented position occurs. Then it becomes clear that there is no “default player” to emulate. There is no such thing as the emulator emulating itself because the emulator is not a player. Bummer.