Leela Zero( A Neural Network engine similar to Alpha Zero)

So in addition to being a neural net, it sounds like a genetic algorithm -- except there is no "mutation" operation?  Or am I misunderstanding the GA concept here?

Well, it is called regression test whether the newly tuned/adjusted idea/pattern is better than the original one.

It is the same in every chess engine tests. 

For example, in SF 9, you have evaluation for Bishop 3.25 and Kt 3.25. And then you make minor evaluation change in new patch where Bishop 3.35 and Kt 3.25. And then there will be regression test between original SF 9 vs newer patch whether your new idea get better outcome or not. If not your new idea will be rejected. If your new idea/patch is passed, it will be accepted for newer development version. There are thousands of similar patches between individual stockfish versions. And each patch have been tested minimum of 10,000 games to prevent statistical flukes. 

A knight sacrifice of Leela. (1sec/move for Fruit vs 10 sec/move for Leela)

NN don’t do any calculation at all (at least if you mean tree search with calculation), but they are used for selecting candidate moves and evaluation. There is a third conclusion of your example: the NN must become stronger. But it is still possible that it plays suboptimal endgame moves or solutions of a checkmate problem.

You can see a similar result in local, tactical Go positions. LZ for example is very strong, but some weeks ago it still played blunder and lost won games by killing its own groups or “forgot” to bring them to live. Or it found the correct solution only after a lot of playouts.

But you are right, NN are better to make strategic or intuitive decisions. Therefore it was doubted that a strong chess engine can be created by using NN, and for me this was the most astonishing about AZ.

I think we will have unseen checkmates and bad endgame moves for a long time in LCZ (and most likely some bug reports about that  ).

By the way: The number of visits was doubled from 800 to 1600, so hopefully we will see some progress.

Is there a reason why no larger network is used? Currently self-play games can be played very fast, even after doubling the visits, so the slowdown of a larger NN should be no problem.

I can still beat Leela (in normal mode - haven't tried hard mode) when playing fast. The trick is to play black and go for my pet line. 

They have plan to increase network. The larger network (i think) will slow down the game play.  They are checking how much max elo can be squeezed out in this current small network.

Tried vs hard, lose one. Leela vs me  , 1-0. But wtf is that she think she can win 79% from black, haha. Dammed it, I analysed with stockfish for any possibility I can exploit her in the opening line I played vs her. She played without any inaccuracy. Very upset. Hard level is  beyond my level or 2200+.

At move 9. I played cxd5 vs her. Stockfish analysis suggest me to play 9.Qb3. As usual, computer evaluations come from deep level of depth 30+. So I deeply followed to the end principal variation of SF and other possible 20 lines why SF suggested me Qb3.

Here is one of my analysis with SF. 

Conclusion, SF suggestion of 9.Qb3 is refuted, there is no advantage in depth 30. SF cant find any possibility to force win/force advantage as white.

Still won at hard level, 1-1(i lose as well). My gut feeling is that hard and normal is not that much different in performance. (may be elo 100-150). 

I don't have an official rating (all I have are these chess.com ratings), but I managed to beat it in normal mode (ID 102).

https://www.chess.com/analysis-board-editor?diagram_id=3939842

<iframe src="https://lichess.org/embed/HrkPImfU?theme=auto&bg=auto"
width=600 height=397 frameborder=0></iframe>

Here is my game in normal mode, bullet game , moves were played fast , and i scored perfect game

https://lichess.org/HrkPImfU

As of now, she is strong tactically, very strong , but still easy to beat on quiet podition and kingside pawn attack.

Wow checkmated her in 20 moves or so.

Normal mode

Played vs Hard mode today in ID 110 for two games.

1 draw simple endgame end as lose.

1 complicated upper hand game also end as lose. Very bad day 

Taking back what i said about tactics, in normal mode she miss obvios one move tactics, and from 60% goes 36%..

Very interesting to follow her progress every day or so.. you can see how she changes openings and what she thinks is good variation.. had fun confusing her and trying to trick her with openings like the grob...very interesting moves she made, one can learn a lot... it will improve my openings as well, she does play like a human, not just brutal calculation that nobody understands.