Too difficult for computers

Yeah, it's not very informative to let it analyse after you have told it the answer, just like if you did the same to a human. It uses what you tell it as part of the analysis, which helps enormously.

But the m41 appeared when I was first entering the moves, on its own, not from going back through the moves.

It will use whatever you have put in during the current session, no more.

Here's a simple endgame for humans to understand but stockfish has trouble with. However stockfish sometimes still gives the correct moves despite not declaring mates throughout:

"Simple" is probably not the word! "Possible" would be more appropriate!

Humans think in a different way. One key is we think of all the squares the knight can reach rather than all the positions that can be reached. The computer can only deal with positions where the two king positions and the knight position are specified. But that should not be too bad for it - without a pawn moving that is merely tens of thousands of positions. Part of the problem is the weakness of Stockfish at deep sacrifices. AIs may be better for this (while weaker at tactics). It's low priority for a key move that will be a problem. Here, it has to look at one sacrifice (in many king positions) work out that accepting it works out and then find the route to the second sacrifice and calculate the results after an acceptance or decline (semantically interesting that the noun for the declining of a sacrifice seems kind of odd).

Computers also seem to struggle in closed positions.

Chess.com's new onboard engine Torch solved #109 in about 5 minutes! After resetting my computer which supposedly emptied the cash. But it is always possible that an engine keep a backup of the most recent analyses on the server!

Well actually stockfish gave most of the right moves, but just didn't declare it as immediately winning. The delay seems to be in the evaluation of the position even if it suggests the best moves or some of the best moves.

That's only natural considering it is not a mate-in-n puzzle. Torch too, only reported a +4.6 score but that would probably indicate it saw the win as it started out with only +1.4 to reflect white's material superiority.

Yeah but if the mate would come within a small number of moves after, such after a pawn breaking through, it usually just adds that to the m long before. For example when I had stockfish play against itself in a queen vs Rook endgame, it gave m20+ long before the Rook was captured (since it is always a max of m10 with a queen vs lone king).

Not sure I understand your point. Anyway, I finished the analysis with just one display line and I screened it to the point where the knight reached the critical square f5. All very trustworthy!

Ah, I see your point. You are still on the trail of delivering mathematical proof which requires delivering checkmate - while the criteria for accepting "wins" in the composition environment are much softer. They will accept not just mates but tablebase endings, endgame theory and high engine evaluation ratings (+3 is sufficient) as satisfactory evidence. Plus evals by human experts. Of course, when you are testing an engine - especially a new one like Torch - you'd need a calibration system to verify the validity of its evaluation scores. Otherwise it's just Torch saying that Torch is winning! Or as we say in Holland "Toilet Duck recommends Toilet Duck" in the category of funniest (real) commercials ever.

The composition community is happy to take chances with top-grade evidence like medical science on the basis that "time will tell" and no one will die from an incidental lapse in the eval systems

Tablebase/endgame theory (should be basically same thing) is fine.

+3 evaluation is inadequate evidence for a good problem, IMHO. An AI would likely leave this significantly short of 100% chance of the result in many cases. Sure, you'll mostly be right. And sometimes you'll be wrong.

Thats a really difficult one

In compositions there is an essential difference between "studies" and "problems". The latter do require proof though it's clear that such is not always attained - especially with complex modern constructions. The judgement of the former accepts a concoction of evidence which falls short of proof. That is why so many endgames were found cooked and seriously dualed when the engines rushed in - not just on tablebases. I found a number of flaws myself.

The idea of the +3 score comes in for the "loose ends" often diagnosed and selected by human experts and the composer himself. Commonly quiet endgame positions (e.g. a rook with 2 pawns against a well-placed knight with 3 pawns) which are near the draw/win divide. Many endgame studies have such endings in or besides the main lines which require reliable evaluations - often to eliminate unwanted duals. In previous centuries composers provided many pages of analysis to convince judges of their viewpoint. Today they'd rather hand that job to a strong engine or by cooperating with one to explore the darkest corners.

But the ultimate defense is in the "reversal of the burden of proof". After engines and human testers have done their job, the endgames are published and even prized and still everyone is invited till the end of times to apply all their resources to find the glitches that were missed. No verdict is final until some electronic beast in some future has confirmed or refuted it by some full proof analytical method.

Interestingly I am viewed as a composer of complex endgame studies though not many. Still today I am fearful that one of those might fall under the sword of an eager kid with a supercomputer. So sad

That was amazing.

Forgive me if I posted this before, but it may have been a different thread. It's called the "kings grave" puzzle:

The computer has trouble seeing this one until about 8 moves away. I know not all the moves are forced, but the queen can always easily reposition to force the king to the a file in every case (too many to include every variation, but have checked them with stockfish, it's a forced mate even if the king runs back to the 3rd rank as he is cut off on the 4th rank anyway)

Delightful artificiality!

What's interesting is that stockfish instantly solves the simpler 12 move version of this Cascade-Underpromotion (below), but can't solve the 19 move one above with the few extra moves interspersed.