The following position is from Dubois – Steinitz, London, 1862. White has just played 8.Bg3.
Smith wrote that none of the engines he tested did very well in this position no matter how long he let them think. In the diagram Black stands much better. His pieces are well coordinated for attack, the P on f2 is pinned, the R on h8 occupies the soon to be opened h-file, the b on c8 will develop at g5, the c6N can easily join the attack with …Nc4, and finally, the f6N can, after …Ng4 attack both f2 and h7.
From White’s point of view his pieces are developed to attack f7, not defend his K. The d3P cuts off the B’s possibility of being used for defense, the g3B has no moves and White’s Q-side is undeveloped. All these indicate that there might be something in the position for Black. Steinitz played 8…h5 and conducted a brilliant winning attack.
So, how did the engines do? After 20 minutes of analysis, Houdini suggested 8…Na5 and after 40 minutes was still suggesting that the position evaluation was = (0.06).
Fritz 12 had a hard time making up its mind between 8…a6, 8…Na5 and 8…Bg4 (= 0.04 approximately).
As Smith points out, the reason the engines don’t find 8…h5 is because all the lines go very deep and while there are few variations, there are few checks which will extend their horizon. The result is that engines are unable to ‘see’ that Black ultimately gains a winning attack with 8…h5. When I made the move 8…h5 I got a red warning light indicating that there was a significant drop in the evaluation. Its immediate response was to show the position as being about 1/3 of a P in White’s favor while Fritz was even more optimistic, thinking White’s advantage was 2/3’s of a P. In the game Dubois declined the P-sac and played Fritz’ move of 9.h4 (0.62) while Houdini recommended 9.b4 (0.16), and 9.h4 (0.13) was its second choice.
This is the weakness of chess engines. Even today, they do not evaluate some positions very well. This is true even in a tactical situation like we see here because the merits of Black’s P-sac don’t show up for many moves which lie beyond the engine’s horizon and it explains why CC GM’s usually end up beating players who rely solely on engine moves.
In the following position where I was White in a postal game played in 2004 I sacrificed a N with 21.Nxf7 without much thought because I was certain that the result would be a winning attack for me. After the game when analyzing it with Fritz 5, it said I was 1-2/3 of a P ahead after 21. Nge4 and its evaluation dropped to only one P after 21.Nxf7.
Today Fritz 12 thinks both moves rate about a one P advantage for White, but Houdini suggests that after 21.Nge4 I’m only a half P ahead while after its first choice, 21.Nxf7, I’m 1-2/3 P ahead.
As Smith points out, some long forcing lines can be calculated by humans better than engines because the end result is beyond its horizon. However, when calculating these variations, it is always wise to do so with the aid of an engine in order to avoid tactical mishaps!
I guess my whole point, two points actually, are: First, Never rely solely on an engine’s numerical evaluation and secondly, studying annotated GM games is still one of the best ways to absorb information. Using an engine helps because it allows you to try out different moves and see why they may not be good. The problem comes in trying to use engines to examine our own games as a means of improvement. Engines will tell us when we’ve made a tactical error, but what they won’t tell us is, what is the correct strategy and why. For that you need books. Books on openings, strategy, tactics, endings and GM game collections. Unfortunately, even then most of us will have no idea why a move is good or bad and will unable to appreciate the things that go into successful planning. Read over Smith’s explanation of the Dubois – Steinitz position and it makes sense. Most of us would never figure it out on our own if we got that position though.