The waiting is soon over 😉, on Wed the announcement will come out, for Friday to start off!

But watch what the machine says – Mate in 19!

For a chess program it simple doesn't matter (if not told otherwise from the programmer), when it checkmates as long as it does checkmate at all (the result is the same anyway).

I had at least ten minutes of this incredible estimation before my eyes.

There are many explanations for this behavior, which is neither a bug nor a glitch. It has been noticed and reported by many other users.

The simplest one is that the engine applies aggressive prunning, which makes it discard candidate moves, especially if it has already hit a promising branch. Another reason is that the engine was not given enough time to iterate while deepening. In the case of Lichess cloud analysis, Stockfish is set up that way, because we are talking about shared and limited computing resources. It is easy to see that this is what happened, as the analysis is part of the Lichess data associated with the game and is seen to "improve" the evaluation once it is fed with the next position.

This is not something special. Engines whose evaluation function consists of an entire neural network, like the old versions of Leela, used to produce greater oddities. For example, when Leela got a very winning position, her conversion method was to skip the most efficient mates and proceed to "give away" all the pieces until she had only a queen or a rook left, and with that she would checkmate. The human commentators interpreted this as "Leela's willingness to torture her opponent", which is poetic, but had a more technical explanation.

I’ve made screenshots to prove the point of fake ‘AI’…

Good. There are many reports like this on the internet and on the Lichess forums. This does not prove that it is a "fake AI". That sounds like a non-sequitor (or is it satire?).

Any kid starting lessons in chess could do better than Stockfish labeled ‘Artificial Intelligence’.

Haha, very funny. It's very hard to generalize with that, as you are cherrypicking something that is not even a bug or a bad solution, technically. The right thing to do is to compare the performance of kid starting lesson with Stockfish in a wide variety of games and positions. We could use a SPRT test, but I think it won't take tens of thousands of games before we can conclude that there is an abysmal difference in elo between these two players.

That is, you have an expert system which collapses as soon as it come to the position which is not in its database…

It's not clear that Stockfish is an "expert system". I also don't understand the role that the "database" plays in your observation.

Thanks for sharing this anecdote.

Thanks for the mention! Funny idea, yes, chess is more fun without AI. Doesn't mean there are not huge databases on how to trick people (into war), are there? While I'm still hoping for some kind of correspondence tourney and simply don't know how something like that might be fun (most favorably chess 960 correspondence). !invest_vote

@anli denkt du hast ein Vote durch @investinthefutur verdient!
@anli thinks you have earned a vote of @investinthefutur !

I don't have a good comment But I enjoyed reading the thread:)...