
The Lichess Game Compressor's Analysis of Game 1
Games 1 & 2 of the Chess World Championships are in the books. The scores, grandmaster commentary, and Stockfish evaluations tell most of the story, but isn't there someone you forgot to ask? That's right, the Lichess chess game compression model has some key insights for us.Trained on millions of games, the Lichess compression model converts moves to bits depending on how good they are. Pretty sure it can give Stockfish a run for its money. Let's see how accurate Gukesh and Ding were in game 1.
Gukesh agreed with the Lichess Compressor's best move 6 times out of 42. Ding, on the other hand, agreed just twice. A sad showing. On average, they both chose the ~11th best moves with medians of 10.5 and 8, respectively. Is this what you would expect from potential World Champions?
In particular, the Lichess compressor noted one massive blunder by Ding: 18. ... Nb2??. This is the 47th best move out of 49 available moves in the position. Insanity. The Lichess compressor recommends the far better Nxe5!!.
From an entropy standpoint we see a clear difference in play. The Lichess compressor uses an average of ~4.4 bits per move for encoding. Gukesh and Ding required an average of 4.98 and 5.21 bits, respectively, indicating that they deviated significantly from established thought.
Gukesh required 209 bits to encode his moves while Ding required a whopping 219! Of that, an enormous 14 bits were needed for his catastrophic Nb2?? blunder. That's more than 6% of the encoding!
We should all pray for Lichess and whoever is covering their storage costs if this is the future of chess.
On an unrelated note, I once made a Lichess bot that used the compression function as an evaluation function. It didn't win a single game.