Page 7 of 8

Re: "Fat fritz" stronger than Stockfish ?

Posted: Tue Oct 29, 2019 11:06 pm
by Albert Silver
kramnik wrote:
Tue Oct 29, 2019 10:44 pm
You just can't put anything on the same plane and for the most of things you are wrong, people is not so ignorant to not get informed on what they are spending. funny that alcoholic aneddoct, thought :lol:
I am also unsure about the argument of price always being proportionate to the work, but regardless will say that thousands of man hours have gone into this project from me alone, nevermind the machines bought and running 24/7 to make it possible. Several new techniques and methods were developed since following the strategies by AlphaZero and Leela would mean being confined to the same requirements as well, notably both required 44 and 88 million games respectively to reach their standard of excellence. At last count Fat Fritz had trained on fewer than 9 million games. These ideas were finetuned and tested over months and months with an impressive number of complete restarts. This net alone is the seventh in the line. Whether people find it a worthwhile purchase or not is up to them, but more than a year of my life has gone into it.

Re: "Fat fritz" stronger than Stockfish ?

Posted: Tue Oct 29, 2019 11:23 pm
by kramnik
You're not alone, buddy. Honest people works honestly with prices. I've always found myself satisfied with the Chessbase products (except the Playchess free app for iPhone but in the Apple universe almost anything sucks). But you should have given to it a more decent name at least XD... "Fat Fritz"... Whatever, the important thing is that the engine is strong.

In my opinion Chessbase should now buy powerful servers and make remote the use of the last generation engine, so there's no way the program can be cracked. In this way people would be forced to buy the software to use the engine. This is an era where to provide digital services more than digital products.
Albert Silver wrote:
Tue Oct 29, 2019 11:06 pm
I am also unsure about the argument of price always being proportionate to the work, but regardless will say that thousands of man hours have gone into this project from me alone, nevermind the machines bought and running 24/7 to make it possible. Several new techniques and methods were developed since following the strategies by AlphaZero and Leela would mean being confined to the same requirements as well, notably both required 44 and 88 million games respectively to reach their standard of excellence. At last count Fat Fritz had trained on fewer than 9 million games. These ideas were finetuned and tested over months and months with an impressive number of complete restarts. This net alone is the seventh in the line. Whether people find it a worthwhile purchase or not is up to them, but more than a year of my life has gone into it.

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 12:13 am
by Nelson Hernandez
Hey Albert, what I want to know is this: is "Fat Fritz" stronger than Stockfish?

Your training methodology sounds an order of magnitude more efficient than anything that has hitherto been devised, but if Fat Fritz's net strength isn't equal to 42850 (Lc0's T40 net) then it may mean that you undertrained your net. On the other hand if Fat Fritz is measurably better than anything out there then you may deserve a Nobel Prize--in that case your training innovations might have important AI cross-applications.

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 2:11 am
by jhellis3
It isn't like 42850 is a very high bar.... I'm sure the Chinese already have private nets +200 or 300 Elo....

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 2:54 am
by MikeB
dannyb wrote:
Mon Oct 14, 2019 6:03 am
Paloma wrote:
Sat Oct 12, 2019 6:29 pm
Found on http://vlasak.biz/evcnews.htm
Eine große Überraschung ist der Motor, bei dem es sich um Fat Fritz handeln wird, modifizierte LC0.
ChessBase tötet mehrere Fliegen mit einem Schuss.
- Er wird Vasek Rajlich los, der den Motor nicht entwickelt, sondern nur seinen Namen wieder verkauft hat.
- Gibt TCEC eine echte Chance, in die Premium Division aufzusteigen.
- Es wird sich großartig verkaufen.

Wenn der Benutzer keine teuren Grafikkarten hat, ist die Engine auf einem herkömmlichen Prozessor leider sehr schwach.
Über die Alternative ist noch nichts geschrieben.

Neueren Trends zufolge wird nur Windows 10 unterstützt.
:lol: :wink:
So a gpu is needed to run Fritz 17? is it confirmed? if that's the case I think I'll skip it for now. In my opinion it's still too soon to sell gpu-only engines
Fritz 17 will be a standard CPU engine, Fat Fritz is is included as a "free add on" to Fritz 17 and will require a GPU to run at home, but you will have access via the Chess Play server. I believe my upstanding is correct - Albert can correct me if I am mistaken.

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 3:54 am
by Albert Silver
MikeB wrote:
Wed Oct 30, 2019 2:54 am
dannyb wrote:
Mon Oct 14, 2019 6:03 am
Paloma wrote:
Sat Oct 12, 2019 6:29 pm
Found on http://vlasak.biz/evcnews.htm
Eine große Überraschung ist der Motor, bei dem es sich um Fat Fritz handeln wird, modifizierte LC0.
ChessBase tötet mehrere Fliegen mit einem Schuss.
- Er wird Vasek Rajlich los, der den Motor nicht entwickelt, sondern nur seinen Namen wieder verkauft hat.
- Gibt TCEC eine echte Chance, in die Premium Division aufzusteigen.
- Es wird sich großartig verkaufen.

Wenn der Benutzer keine teuren Grafikkarten hat, ist die Engine auf einem herkömmlichen Prozessor leider sehr schwach.
Über die Alternative ist noch nichts geschrieben.

Neueren Trends zufolge wird nur Windows 10 unterstützt.
:lol: :wink:
So a gpu is needed to run Fritz 17? is it confirmed? if that's the case I think I'll skip it for now. In my opinion it's still too soon to sell gpu-only engines
Fritz 17 will be a standard CPU engine, Fat Fritz is is included as a "free add on" to Fritz 17 and will require a GPU to run at home, but you will have access via the Chess Play server. I believe my upstanding is correct - Albert can correct me if I am mistaken.
Free add-on isn't quite accurate, since it really is the showcase piece of Fritz 17, but yes, Fritz 17 will actually come with several engines. There is a new Fritz 17 engine, a conventional AB engine that has no relation to Rybka or Vasik Rajlich, so no idea why his name is dragged into this at all. There is also Fat Fritz, my neural network that runs in a slightly modified lc0. There will also be a smaller Fat Fritz Jr. network that should run well on weak GPUs and even half decently on a CPU at longer time controls. Finally, it also comes with Stockfish (just as in Fritz 16) and Leela itself. Leela will come with the 42850 network, automatically installed and configured. You would not believe how many people have requested this last point.

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 4:16 am
by dannyb
Albert Silver wrote:
Wed Oct 30, 2019 3:54 am
There is a new Fritz 17 engine, a conventional AB engine that has no relation to Rybka or Vasik Rajlich, so no idea why his name is dragged into this at all.
Because on this page (https://www.niggemann.com/en/fritz-17-deutsch.html) - Details - it says he is the author of Fritz 17.

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 4:53 am
by Albert Silver
dannyb wrote:
Wed Oct 30, 2019 4:16 am
Albert Silver wrote:
Wed Oct 30, 2019 3:54 am
There is a new Fritz 17 engine, a conventional AB engine that has no relation to Rybka or Vasik Rajlich, so no idea why his name is dragged into this at all.
Because on this page (https://www.niggemann.com/en/fritz-17-deutsch.html) - Details - it says he is the author of Fritz 17.
The description is indeed the official text, but the 'details' are from the store and just wrong.

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 5:13 am
by kramnik
Albert Silver wrote:
Wed Oct 30, 2019 3:54 am
There is a new Fritz 17 engine, a conventional AB engine that has no relation to Rybka or Vasik Rajlich, so no idea why his name is dragged into this at all. There is also Fat Fritz, my neural network that runs in a slightly modified lc0. There will also be a smaller Fat Fritz Jr. network that should run well on weak GPUs and even half decently on a CPU at longer time controls. Finally, it also comes with Stockfish (just as in Fritz 16) and Leela itself. Leela will come with the 42850 network, automatically installed and configured. You would not believe how many people have requested this last point.
Quite interesting!

Re: "Fat fritz" stronger than Stockfish ?

Posted: Wed Oct 30, 2019 10:08 am
by crem
Albert Silver wrote:
Tue Oct 29, 2019 11:06 pm

I am also unsure about the argument of price always being proportionate to the work, but regardless will say that thousands of man hours have gone into this project from me alone
Few thousand of man hours is probably more than all human effort in LCZero project combined!
One full-time working year is typically 2000 hours.