An important further point is that the probability of a turnaround is actually much smaller still than your estimate, because giving material odds from the starting position (1) is known to cost more elo as you rise in elo, and we are talking about perfect play here, and (2) because Stockfish evaluates us as a Rook up in an early endgame not opening. That's a much more decisive position to be in.

@Grade'Eh'Bacon just because you provided a good answer doesn't make it a great question - you answered the question 'what is the relationship between eval and winning probability', an altogether different and deeper question than what the user is asking, 'can Black hold this position with perfect play'.

A typically erudite answer from you on this topic! Can I ask, why consider the second puzzle you show 'cooked in 60 moves', as opposed to simply a valid s#60? (Such that the record for dualed selfmates would stand at 60...(?))

@David Do you have any evidence that he "had lost a game of chess a couple of days before making that quote" or just baseless conjecture? Had you bothered to read Wells' essay, you'd have known the quote is literary, his feelings towards chess aren't remotely negative - quite the opposite - and that Dunning-Krüger has no role here. Instead what you wrote could apply with equal probability to anyone who has a negative opinion about chess, and could have been saved by a Google search (billwall.phpwebhosting.com/articles/hgwells.htm).

@user21820 I want to commend you for an interesting idea that provides food for thought! But, I'm not sure it's so different from the above (choosing the nth best move): evaluation assumes optimal follow-up, so the introduced weakness can only be exploited via near-optimal play, while for a player of low rating the position may not appear any worse. (By "restrict to fixed time per move", I don't see what difference that would make to the moves played: guess you refer to time taken.) Btw how about "it is definitely possible for a computer to simulate human players" - what method do you imagine?

The question asks about the retraction. Your answer talks about the original post

@user21820 Let's dig a bit further to see how your idea could work meaningfully. By "each move", do you mean 'on each move number' or 'in each given position'? If the latter, don't you have the problem of generalizability to unseen positions, and if the former, wouldn't strange play remain in the game since the computer must incur a certain amount of loss at certain times?

"Understanding" his moves isn't the same thing as predicting or imitating them, however.

@firtydank At the risk of questioning someone who may have really looked into it - is there real sustained evidence that Houska prefers to be identified as a WGM? I watched the FTX Crypto Cup 3 broadcast's beginning and saw she was introduced as WGM there, but has the same happened on any other day? (For example I didn't find any introduction on Day 2. It's possible technical organizers may have just slipped.) If so I find it very surprising and an interesting spot. But I'd like to see this evidence.

@RemcoGerlich gonna disagree with you there - sure rating differences may make a much bigger difference to moves played than personality/playstyle, but playstyle certainly affects what ideas you go for and is an interesting problem for chess engine research as I understand it

Can you provide some details? How did they try to imitate famous players? I'm not aware that Rebel made any effort to do this and ShashChess almost certainly didn't (and it wasn't an engine 'back in the days')

What do you mean by validation? If you mean cheat detection, this is probably on-topic. If you mean legal move detection, it's a coding problem (and seems a trivial one: just generate the legal moves and check the proposed move is in that set...)

See my answer below: as of analysis conducted in 2021 the mystery is all but solved.

Nope. Chessbase is talking about the article Short published his original statement in, not the retraction

@user21820 What do you mean by "centipawn loss distribution"? Distribution as a function of what? NB: People have tried experiments like Stockfish picking a lower option (but still high) than its top choice unless this costs a lot of eval. It didn't work, the engine is still invincible to humans. (Why? Because in most classical games, and almost all faster games, humans lose because of significantly weakening moves, not slightly suboptimal choices.)