The bottom line is the more calculation and the less evaluation position requires -- the better chances you have.

I thought I would never read something like this in my life. Are we at a position where AI's weak points are calculation and not evaluation? This is the opposite of what everyone considered to be difficult for AIs in Go.

I thought I would never read something like this in my life. Are we at a position where AI's weak points are calculation and not evaluation? This is the opposite of what everyone considered to be difficult for AIs in Go.

This is pretty cool. It makes more sense once you understand what actually goes into the AI. The big breakthrough for recent computer Go has been finding how to create a neural network to evaluate the board and provide whole-board positional judgment, and so that's precisely the area where they're relatively strong now. You can think of neural networks of this sort as being the computer analogue to human pattern recognition, instantaneous feeling, snap judgment, etc (an important part of what Kahneman's calls "system 1", if you are familiar with that literature about human cognition). In other fields, these neural nets have provided huge advances in image and facial recognition, speech processing, etc.

Over millions of training games, more data than any human could possibly process, the "value net" is able to refine its overall global judgment to be much better than any human's, simply by learning over that massive amount of data exactly how much to value different kinds of influence and territory and thickness, and a finely tuned instinct for when a result is better or worse based on the whole board situation, so they outperform humans in the opening and in overall judgment.

But the value net doesn't provide quite as much help in tactical fights and trying to kill/live with a large dragon. To reiterate, the value net's role is most analogous to human snap judgment about what results are good and bad and overall board feeling. Complex tactics involving large dragons aren't so easily reduced to simple snap judgment and overall feeling - it requires exact reading to check whether you actually can get the second eye or connect to something or not. While snap judgment, pattern recognition, etc, can still help spot possible moves that might be tactically useful, and judge whether living might be more or less likely, in the end you still have to do the reading. While bots aren't exactly weak at reading any more and are still steadily improving, it's not the primarily the aspect of the game improved by the latest breakthrough.