Stockfish is blind!

Don't bring this forum back please

I think Bd2 also wins.

Current Stockfish 17 at 4000 elos is much harder to beat.

Anyone having achieved that feat?

How frequently?

Anyone giving Stockfish odds of a knight or so?

Just asking.

Anyone giving Stockfish 17 full strength the odds of a knight and beating it?

Post games.

After 1. Qf6+ Bxf6 2. gxf6+ Kg8 only 3. Bd2 wins convincingly, Ng5 is weaker.

By the way, the position is so complicated, like most closed positions, that even latest Stockfish 17 is UNABLE to see the correct continuation and win. It indeed finds the queen sacrifice fast, wasn't the case with previous versions, however goes astray with 3. Ng5 instead of 3. Bd2 slide, aiming to reach h6, and then stumbles onto a white advantage, but not full win.

At least with faster check.

So the position is astounding indeed.

Don't laugh.

Is the 5 consecutive posts limit still in place?

If so, someone please answer briefly, so that I can add more comments.

May I answer in a non-briefly manner?

I tried the position using Lc0 v0.32.0 + 'BT4-1740.pb.gz' (1024x15 NN) - finds the path in less than one minute running on CPU (openblas, as I have no GPU) and only two cores from my i5 laptop (resulting in two nodes per second):

0:01: -0.60/1 1.Nd2 Nxa30:02: -0.38/2 1.Nd2 Nxa3 2.bxa30:03: -0.48/2 1.Nd2 Nxa3 2.bxa3 Bxa3
0:06: 0.00/3 1.Nd2 Nxa3 2.bxa3 Bxa3 3.Qf6+
0:07: 0.00/3 1.Nd2 Nxa3 2.bxa3 Bxa3 3.Qf6+
0:08: 0.00/3 1.Nd2 Nxa3 2.bxa3 Bxa3 3.Qf6+
0:08: +2.37/3 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 Nxa3 4.Rxh7
0:09: +2.02/4 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3
0:10: +1.89/4 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 Nxa3 4.Rxh7 Nc2 5.Bf2
0:13: +1.72/5 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 Nxa3 4.Rxh7 Nc2 5.Bf2 a3
0:18: +1.92/5 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2
0:22: +2.15/6 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2
0:23: +2.15/5 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2
0:25: +2.08/6 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2
0:27: +2.05/6 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2 6.Nf3
0:29: +2.03/6 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2 6.Nf3 a3
0:33: +2.04/6 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2 6.Nf3 a3 7.Bh6
0:38: +2.15/6 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2 6.Nf3 a3 7.Bh6
0:42: +2.09/7 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2 6.Nf3 a3 7.Bh6
0:47: +2.06/7 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 Nxa3 4.Rxh7 Nb5 5.Rg7+ Kh8 6.Bf2 Nxc3
0:52: +2.16/7 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Ng5 h5 4.Bxh5 Nxa3 5.Bd2 Nc2 6.Nf3 a3 7.Bh6
0:58: +2.26/7 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5
0:59: +2.34/7 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4
1:00: +2.40/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4
1:05: +2.47/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4
1:10: +2.65/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4
1:15: +2.69/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4
1:15: +2.66/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4 8.Nxe4
1:20: +2.68/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4 8.Nxe4
1:21: +2.65/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4 8.Nxe4 dxe4
1:26: +2.65/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4 8.Nxe4 dxe4
1:31: +2.53/8 1.Qf6+ Bxf6 2.gxf6+ Kg8 3.Bd2 Nxa3 4.Bh6 Nb5 5.Bg7 Nxc3 6.Ng5 h5 7.Bxh5 Ne4 8.Nxe4 dxe4

Yeah, nice try.

Thanks.

Is that engine a clone of Stockfish, almost all top engines used to be in the past?

I have missed some engine developments in the last 5-8 years or so.

I still remember how I beat myself older Stockfish on talkchess forum from a position that led here, and my friend Louis, running Stockfish on big hardware was astounded to see the queen sacrifice, didn't believe at first, and Stockfish needed hours to even have a clue.

Good old times, not any more.

No it is not; it is the current latest version of the Leela Chess Zero project. Not a NNUE engine like SF and many others, a pure NN powered using MCTS. Performance would request a GPU, yet already interesting to run on a CPU. Home page: https://lczero.org/

Should you like to play a real gambler, select the biggest net (https://lczero.org/play/networks/bestnets/) BT4-spsa-1740 and rename the lc0 exe file with . lc0-valuehead to set the engine to play without any in-depth search (one node per move; moves result from the bare policy head). No need for a GPU, you get instant play anyway - yet at GM level. Worth a try, trust me you will appreciate!

p.s. lc0-policyhead not lc0-valuehead. Sorry for that.

I might try at some point.

Monte Carlo is not much different, you are probing game outcomes one way or another, and you still need evaluation.

I have huge experience following clones.

Every single engine claimed not to be a clone, yet Rybka, Houdini, and many other Stockfish clones used to be.

Also the top engines copied many ideas from free Stockfish code.

That is a fact.

Curiously, as soon as Stockfish implemented knowledge to gain 50-100 elos over them, the copiers instantly caught up.

Wonder? I guess the truth is obvious.

It's extremely difficult to make a 4000 elo engine, almost impossible, and Stockfish use thousands of cores, contributors, main developers etc.

Same was true of AlphaZero, I bet it used some Stockfish ideas.

It was big hype, but current Stockfish is 20 times stronger than AlphaZero.

Google simply used tremendous hardware, and Stockfish was handicapped.

Also current Stockfish plays awesome strategic chess, all AlphaZero knew was to attack wildly, but for that you only need big hardware.

So no, I don't trust the idea every single engine is original.

The top in any case borrowed quite much form open Stockfish.

Or you think they didn't look at the source?

Or didn't get tempted?

You bet.

Limiting search depth, plies is an interesting idea.

I guess I was the first to suggest perfect chess could be played at ply 1, by having perfect evaluation.

That was the purpose of The Secret of Chess.

You mean 2600 strength at ply 1, without search?

I guess they are doing at least quiescent, otherwise you will capture all of their pieces.

But the idea is interesting, might try it in the future.

I doubt it, Fritz, aka Rybka is 3200 elos, Stockfish 4000, huge difference.

Stockfish plays tremendous strategic chess, better than Kasparov, better than Carlsen, better than Fischer.

Rather in the ball park of 2500. I think it is the baseline strength ranking for GMs. Getting GM norms indeed require 2600 performances, if I am correct (under some conditions, of course).

No quiescent search, as far as I understand.

Oh, you missed part of the journey, indeed. Rybka-Fritz was EOY 2015 to EOY 2017 (FRITZ 15, 16 engines authored by V. Rajlich). Since then (close to 8 years, so) the engine is Ginkgo from F. Schneider. Welcome back!

Thanks Kiwimotard, Kiwimaquisard, I mean.

Yes, but Ginkgo was also suspected or proven of copying, if I am not wrong.

That is what I remember remotely, though I haven't followed closely developments in the latest years.

I remember watching Ginkgo games, the style was suspicious, besides the name itself elicits copyright issues.

It's also the gut feeling, Fritz accepted one clone, so why not a second one?

But I might be wrong in the specific case, my point is it is extremely extremely hard to make a top 3500+ engine nowadays, UNLESS you copy from freely available sources.

You know, all the good ideas that work in an engine, might very well work in another too.

At least I was watching closely as most good Stockfish evaluation ideas worked in other engines too, when tried.

It's simply the correct knowledge. And the same would be true of search.

Some ideas might be engine-specific, though, hinging on the way engine parameters are tuned.

8 years ago you claimed that Ng5 is mate in 25. And now it's not convincing and weak?