" End of an Era" is coming?

[WinPooh]

Well, this is end of an era for Houdini and Komodo.
This mean Leela become 99.9% best engine within a year!
Beating SF will become 100%.

After 0-2 vs. S and K this table looks a bit outdated.

[AdminX]

I still don't care for her silly (Trolling) endgames.

[M ANSARI]

I remember coming on talkchess a few months ago and this LC0 was just getting started. It looked promising but I expected at least 3 or 4 years before something that could reasonably compete with the top 3. I am totally amazed that already LC0 can go toe to toe with a very mature and most powerful engine on earth. If progress continues at this pace, I really doubt any other engine will be able to compete with LC0 unless a radical change in how AB engines work. Most likely AB engines will have to start using some sort of hybrid setup that also uses NN.

[duncan]

I remember coming on talkchess a few months ago and this LC0 was just getting started. It looked promising but I expected at least 3 or 4 years before something that could reasonably compete with the top 3.

Indeed exciting times for computer chess. But all lc0 had to do was emulate alpha zero. Why should that need 3 or 4 years?

[Guenther]

I remember coming on talkchess a few months ago and this LC0 was just getting started. It looked promising but I expected at least 3 or 4 years before something that could reasonably compete with the top 3.

Indeed exciting times for computer chess. But all lc0 had to do was emulate alpha zero. Why should that need 3 or 4 years?

Oh boy,

1. until very recently most stuff from A0 was hidden from public (some is still unknown)
2. emulating googles hardware for the training is only possible by a factor 1/X
3. getting enough participants and setting up a framework for collected training
4. time and money
(5. whatever still lacks)

[Milos]

2. emulating googles hardware for the training is only possible by a factor 1/X
3. getting enough participants and setting up a framework for collected training
4. time and money
(5. whatever still lacks)

Google (allegedly) did the training games in 9 hours using 5000 first gen. TPUs which is 45000 TPU hours. First gen. TPU is roughly equivalent to RTX2070.
If you have only 10 participants you can make all training games in only 6 months. 10 participants that is less than 10k US$ in total.
So basically all 3 points of yours are the same BS non-existent issue.

[Guenther]

2. emulating googles hardware for the training is only possible by a factor 1/X
3. getting enough participants and setting up a framework for collected training
4. time and money
(5. whatever still lacks)

Google (allegedly) did the training games in 9 hours using 5000 first gen. TPUs which is 45000 TPU hours. First gen. TPU is roughly equivalent to RTX2070.
If you have only 10 participants you can make all training games in only 6 months. 10 participants that is less than 10k US$ in total.
So basically all 3 points of yours are the same BS non-existent issue.

just ommitting 1, which makes all other points moot, because training doesn't get same efficiency...

and why not just add to 5 your point from the past?
5. needs programmers with professional high expertise in NN and computerschess

you made this point already several times, but obviously you just needed to disagree for fun as usual

[Nay Lin Tun]

Update, Leela is currently no 2 , in both TCEC and CCCC (new nets are still getting stronger, ?minimum +10 elo above cccc net and +30 elo above tcec net)

[AdminX]

Well I think Stockfish will win but it is closing the gap. By the end of the year she just might be number 1.



https://blog.lczero.org/2019/01/leela-p ... .html#more

[M ANSARI]

Not sure who is responsible for LC0, but holy crap ... what an incredible achievement! I would have never thought it would be possible to even be in the top 50 engines a few months back! Also what a monster SF has become! It seems like only recently where SF was slowly climbing the engine ladder and now it has become a real beast!