I had the feeling that your words “many megabytes and millions of operations” were meant to trace a line between computers and humans. I think they’re doing the same work and I think that our brain too has a lot of memory and does a lot of elementary operations.
Of course we are slower and weaker on repetitive tasks. But I think that the intuition for the first next move is something relatively quick and light, both for AI and humans.
Reading is the hard work (that I try to avoid! )
You’re right: I’m making some confusion here.
I’m talking about NN instinct, but what I see on the board when I review my games is the outcome of katago using MC to refine NN results. So that is actually instinct AND reading working together.
Hey, I won a 9 stones handicap game against it!
Another case where points and winrate tell two different stories:
Winrate
“Oh, well, black won this game from the first moves”
Points
“Wow: a handicap game where black nearly lost all his advantage!”
Oh, but funnily enough, this is also what happens when training the Neural Network (= the engine’s “instinct”)!
During training, the reading (MCTS) is used (through “back-propagation”, so the core aspect of modern machine learning) to “correct and hone” the Neural Network.
So I think it makes perfect sense to try to use the AI moves (or the moves of a pro) to hone your own instincts. Well, I don’t necessarily know that it’s an efficient strategy, but it makes sense to me.
I genuinely think this is actually how a lot of the improvement happens at the DDK level: very often, in my impression, the difference between a 12 kyu and a 20 kyu is not better reading abilities as much as the fact that the 12 kyu has learned to “imitate” good players in more situations.
For example, when I look at a situation, I can think “a shoulder hit seems good”, or “making a base with a 2 space jump is the obvious move”, “peeping a tiger’s mouth can give me a free move to build shape”, etc, but there was a time when I had never seen those moves or been told they are good.
Or a basic example: a total beginner often doesn’t know that playing corner moves is the best in the opening. But as soon as they’re told and they notice that good players do it all the time, beginners start playing 4-4 all the time, simply because it’s the one move they’ve learned so far. When they learn about the 3-3 invasion, their instinct goes there all the time, simply because it got trained to think “that’s a good move”.
I remember learning that 5-5 was a good starting move on 9x9 simply by observing what the players beating me were playing. (I was going corners-side-center and playing 3-3 every game, like many other advanced beginners do)
Sure, I wanted to point out that a reliable intuition needs to be backed up by huge amounts of data, both for AIs (166 MB) and for humans (learning many tesujis, josekis, shapes,… and life and death problems). I don’t think it’s possible to become strong at life and death problems just by memorizing solutions. I mean, it’s theoretically possible but the human brain doesn’t work like that, we remember solutions that we understand, and we understand solutions after putting some effort into reading variations.
do you mean : “players misjudge when to attack and when to defence, and, misjudge which groups/stone / move is strong, and which groups // // are weak” ?
)
