Are there any positions where computers would play a bad move, but we humans know of the existence of a better move?
By “computer” I mean standard engines such as Stockfish, at full power.
Edit: By a “bad” move I mean a move that changes the "shape" of the game.
A good example is the following position from a study by William Rudolph (which is also in the posting linked to by ASTA):
3B4/1r2p3/r2p1p2/bkp1P1p1/1p1P1PPp/p1P4P/PPB1K3/8 w - - 0 1
1. Ba4+! Kxa4 2. b3+ Kb5 3. c4+ Kc6 4. d5+ Kd7 5. e6+ Kxd8 6. f5
The problem here is that White seems to be completely lost no matter what they do, but it is in fact possible for them to lock up the position by giving up even more material.
In general, fortresses are still difficult for standard top engines. It is not hard to come up with positions that current top engines do not understand. It is, however, very difficult to achieve such positions in real games with any reliability.
It is also worth noting that there are engines that have specialised fortress-detection code. For instance, the Stockfish fork Crystal solves the position given above in a matter of seconds. However, these engines are weaker than the top engines in normal positions.
