Maybe the last stand in standardchess in which we can beat the computer is to recognize and build fortresses. There exist a lot of definitions for fortresses but in this article I stick to endgames (maximum 4 pieces besides pawns and kings) in which material is sacrificed to defend successfully. Some trivial endgames like rook-pawn + bishop of the wrong colour are evaluated correctly by our current engines but a fortress like in the high-class game Shakhriyar Mamedyarov - Fabiano Caruana created a lot of confusion.
Commentators initially thought that Fabiano blundered and even after the game some players weren't convinced about the fortress, see chessbomb or the blog of James Stripes. Today we are so dependent and addicted to evaluations of our engines that we don't question them anymore. Personally I only believe a score of +5 or a tablebase-hit that a win is 99,99% certain as I described earlier in my article to analyze with engines. Setting +5 is really not too high as some time ago the scores went up high while analyzing an endgame with Stockfish and Komodo while eventually they didn't manage to find the killer.
On the other hand above analysis of both fortresses also show that engines can still discover a lot of interesting lines. Not seldom a strong engine can crack a fortress by using some very complicated constructions. I had a look to the so-called fortress which appeared in the marathon-game of 122 moves in the last world-championship. Neither Carlsen nor the commentators found a win in the 7th game but thanks to Lets Check I still found a weak spot.
While preparing this article, I also reviewed some analysis of fortresses popping up in my older games. No surprise here too as I also found some holes in the analysis. I mentioned in my article the Spanish with d5 that a fortress exists in the analysis of move 50 of my game against Fabien Libiszewski. Stockfish must move heaven and earth but in the end manages to breakthrough.