On February 8, 2024, a new AI model has achieved a significant milestone in the realm of chess, reaching a Lichess blitz Elo rating of 2895. This 270M parameter transformer model was trained with supervised learning on a dataset of 10 million chess games, annotated with approximately 15 billion data points. It has been reported that the model can play at a superhuman level, achieving grandmaster-level chess without the need for search algorithms, domain-specific tweaks, or explicit chess knowledge. The model was distilled from a 3.6k ELO heuristic/search-based expert into an end-to-end model without pretraining, by distilling Q-values and not just oracle moves. It has also been noted that the model generalizes well to unseen board states and can solve challenging chess puzzles. This AI model is an approximation of Stockfish 16, which leads to strong chess play when used in a policy.
On February 10, 1996, IBM's Deep Blue won a landmark game against Garry Kasparov, marking the first time a computer defeated a reigning world champion under standard chess tournament conditions. In the end, Kasparov still beat the machine 4-2. https://t.co/PnRFx5MTCB
“Grandmaster-Level Chess Without Search” New AI insights. Article: https://t.co/vXLIJt8jDq
Grandmaster-Level Chess Without Search https://t.co/tmBi5hvtDk
"it is possible to distill an approximation of Stockfish 16 into a transformer via standard supervised training. The resulting predictor generalizes well to unseen board states, and, when used in a policy, leads to strong chess play (Lichess Elo of 2895 against humans)" Awesome! https://t.co/XrqocJKZsD
Distilling a 3.6k ELO heuristic/search based expert into a 270M end-to-end model (with no pretraining) by distilling Q-values and not just oracle moves by themselves. Nice! Resulting model is still pretty much superhuman at 2895 ELO. https://t.co/yqrverEHUB
Distilling a 3.6k ELO heuristic/search based expert into a 270M end-to-end model (with no pretraining) by distilling Q-values and not just oracle moves by themselves. Nice! Resulting model is still pretty much superhuman at almost 2895 ELO. https://t.co/yqrverEHUB
Distilling a 4.2k ELO heuristic/search based expert into a 270M end-to-end model (with no pretraining) by distilling Q-values and not just oracle moves by themselves. Nice! Resulting model is still pretty much superhuman at almost 2895 ELO. https://t.co/yqrverEHUB
Transformers can now play chess! "We train a 270M parameter transformer model with supervised learning on a dataset of 10 million chess games. We annotate each board in the dataset with ... roughly 15 billion data points. Our largest model reaches a Lichess blitz Elo of 2895". https://t.co/9wuNaM0A34
Google Deepmind presents Grandmaster-Level Chess Without Search paper page: https://t.co/Drw8kduoYH largest model reaches a Lichess blitz Elo of 2895 against humans, and successfully solves a series of challenging chess puzzles, without any domain-specific tweaks or explicit… https://t.co/vbqUockfqQ