http://www.mediafire.com/file/s99iadlt6d6l93e/LCZero+Gen10.7z
This is current LeelaZero files, generation 10,
1. download and unzip files
2. double click play
3. double click setup--->setup your username and password
It will run your user name in Leela zero server http://162.217.248.187/ .
One game usually takes a few minutes, and you can review every game when they finish.
This is my user name, and the games will be able to reviewed immediately by clicking game ID ,(powered by chesstempo GUI. )
Leelazero is learning very fast, he improved too quickly in a few days and he is now comparable to stockfish level 3 ( rating around 1200-1300).
Leela successfully learned "Windmill drill" to opponent Queen by his own
(check move 17 to 24)
Me vs Leela Zero, 1-1
Leela Zero becomes much stronger than yesterday and he is competing well againt me in 5 mins blitz today, I have to play seriously to win him. ( I kept pushing in last 20 secs without knowing that leela resigned already. )lol . My guess is that Leela is at least 1600-1800 rating now!
At move 15. Leela did not expect the tactical shot 16. Rxe7! I did not see that tactical shot for a while and I was panic for but i saw it.
Leela is obviously strong enough to play against average human now. @DeirdeSkye . We are very excited in testing Leela. Leela already had 570k training games and reached 1600 rating I guess. Alpha Zero had 44 millions training games (9 hours) and beat SF 8.
Leela Zero has much lower layers of NN but the programmers think that they have better NN structure and Leela Zero might not need that 44 millions games to surpass the stockfish! At the moment Leela is improving very fast within a few days, however, no one knows when the progress of NN learning will be slowing down.
There is still massive 2000 elo gap between Leela's 1500 rating and stockfish's 3500 rating.
Does Leela Zero have a better smositional pattern recognition? And how does the smositional evaluation function work?
Today Leela becomes generation 14, (a game vs me was played vs generation 12). Leela is fearless, aggressive without caring king safety. 
Stockfish vs Leela rating list! 
It's clear that the engine has never seen a position where one side was winning by 3 queens, a rook, a knight, and a bishop. It took 40 moves to mate from that position, when it was just trying all the possibilities.
@Attacker5000, For some reason Leela didnt care M5 or M50 as along as the result end as a win. The same as human, we dont search the shortest line in mate. Probably due to 50 moves rules, where there is no bonus for winning in 5 moves vs 50 moves.
For those peope who are wondering how Leela Zero do search and evaluate the position! Here is the snapshot of Leela running on my Geforce GTX 1060. Leela is searching" depth 16, nodes 800, nps varies (from 1.5k to 17k), evaulation -same cp-value as other engines (1.00 for one pawn), winning probability- 0%-100%.
I agreed to a certain extent! It is the initial working model of Leela , the instructions are given as clearly as possible. There is much difference between " Search a mate within 50 moves vs Search the shortest mate" . If former task was asked, the program will stop searching whenever he see any mate from M1 to M50. If latter task was asked, the program will keep searching even though he see M10, M5, cos his task was incomplete and he dont know which mate is the shrotest. That instruction may sometimes require massive amount of computing power and sometime results in hanging of the program. ( For example, if you ask a program to find a prime number, the program will give you any prime number 2,3,7 etc, however if you ask the program to find the largest prime number- the program will hang cos it will keep searching the largest prime number which is impossible. )
Yes, I believe that Leela will eventually fix on his own. If you look at the first game of leela, he was playing really terribe, but he fixed by his own.
I am afraid, i am not the programmer of Leela(I give credit to all programmers in my first post) , I am just the volunter supporter/tester of Leela. I understand basic programming as amater though.
Not seeing how this any different from SF.
Is it playing itself and learning or people just inputting whatever game they want to and calling it instructive?
Neural network is the same as a learning child. If a child hits the spike, he get the feedback that hurts him. And he will always avoid it in the future. This is the simple input-->output-->recognizing the result--> implementing in the future. ( However we dont know how his inside brain work. ) This link will explain how neural network work. http://www.explainthatstuff.com/introduction-to-neural-networks.html
Stockfish Author Garry Linscott and other stockfish programmers started Neural Network Engine similar to Alpha 0. The project was started since last month , with the idea of successful machine learning from Leela Go Zero and Alpha Zero . Leela Zero started from scratch with random moves (0 elo). I am a bit interested in the project and contributed a few minutes of machine power from my laptop. They said current elo is around 500-1000. Here is the game. If you are interested and would like to contribute your machine power into the project, let me know. I will help you set up your machine.
It is super easy to set up.
Enjoy..