from https://t.co/5dZrvcax5x:
This is actually Black's best move, moving the bishop from c4 to e6 to maintain control and continue Black's winning attack - after this move, Black can force checkmate in 5 moves starting with a rook invasion to c1, leading to a mating sequence where Black's queen and rook coordinate to trap White's king on a8.
Also, the reason for this is because when explaining the best move, the model uses a PV that assumes your opponent is playing at the highest level. We are planning on adding a feature where you can set the level you want your opponent to play at in the simulation, so at lower levels it will explain that the goal is to checkmate because opponent would be playing worse and that can actually happen
Analysis from https://t.co/5dZrvcax5x :
The move e4f6 is an excellent choice as it maintains the mate in 3 threat for White. By moving the knight to f6, White creates a powerful attack on Black's king, threatening the h7 pawn and controlling key squares around the black king. The principle variation shows that Black's best response is to capture the knight with the g7 pawn, but this weakens Black's king position further, allowing White to continue the attack with Bd3 and eventually Bh6, leading to a forced checkmate sequence.
Grateful on this Thanksgiving ๐ฆ to be a part of @grok's api beta test. Attached is a board position, and here are Grok and Sonnet 3.5's analyses of it:
Claude Sonnet 3.5 response:
"The move d4c5 (pawn captures pawn on c5) is a decent move but not the best, as it allows Black to develop their knight to c6 and then capture the c5 pawn with their bishop. White responds by preparing to develop their queen to g4 and knight to f3, putting pressure on Black's kingside, but this comes at the cost of losing a central pawn. For beginners, it's important to note that while pawn captures can be tempting, they sometimes lead to loss of tempo or material if not carefully considered."
Grok response:
"The move d4c5 (capturing the pawn on c5 with the pawn on d4) slightly decreases the position's evaluation from 55 centipawns to 39 centipawns, indicating it's not as strong as the best move, but still reasonable. After this move, Black responds with Nc6, which attacks the pawn on e5, prompting White to play a3 to potentially support the bishop's development or prevent a knight move to b4, showing that the initial move d4c5 didn't disrupt Black's plan significantly but did offer White some positional control."
๐ First look at our latest ChessPredict feature!
Scan any chess position โก๏ธ get an interactive board โก๏ธ analyze your moves in real-time ๐ง
More coming:
Full variation analysis
Loading bars
Audio controls
Try it & tell us what you think! ๐ค
#Chess#ChessTech
Now another interesting thing - Gukesh's response to Ding's unorthodox move was optimal! Well, why?
Let's see what ChessPredict says.
It looks like Gukesh wants to maintain pressure on the center, fortifying his protection of the e5 pawn, maybe taking advantage of the fact that Ding is not advancing for more control of the center ๐ค
In the first game of the world chess championship, we used https://t.co/5dZrvcax5x to analyze Gukesh and Ding's moves. Particularly an unorthodox move by Ding that left everyone suprised. Was this a non-standard move in this situation? Let's ask ChessPredict, powered by stockfish!
It looks like the move was indeed unorthodox, but Chess Predict's analysis explains that advancing the queenside pawns is a strong tactic, so Ding's unorthodox move of a7a5 still achieves this goal, and is a protected pawn by the knight, although he is not advancing the center as much as b7b5 would.
This leads me to a new ChessPredict feature that is in the works! A simulation feature, where we can select what piece to move in a given board situation, and see a holistic evaluation of the move in question, explaining the sequence that follows. Stay tuned!
people seem to be hating on this
but itโs not like weโve seen any error bars at all so far in any evals published by any major provider??? is this not a good thing???
There's a lot of work to be done on chess predict - including improving our vision model to handle more diverse distribution of chess boards, if you want to get involved DM!
Chess Predict is participating in the google FIDE kaggle competition on creating an efficient AI chess bot!
After the competition ends (3 months from now), we will post a comprehensive thread explaining what we did and how we used our own extension to improve our bot!