Hello, namespace, you wrote: N> N> Unlike the previous programs DeepMind, AlphaZero formed as the algorithm, capable to learn at once to several tasks-games, instead of one. For this purpose algorithm did not train to win, and gave only basic knowledge of game rules. AlphaZero then played with itself and independently worked out tactics. N> I transport on language: They rules, and then launched search to find all successful variants. N> AI? N> you are valid facepalm certainly in course what to sort out all variants in chess up to the end - it is impossible, and there is a problem of an objective estimation " positions" so that in absence of the unambiguous definiteness to make their most exact estimation? And so, AlphaZero searches just 80 thousand positions per second in chess and 40 thousand in shogi, compared to 70 million for Stockfish and 35 million for Elmo. I.e. Having computing productivity of depth of courses almost in 1000 times less (at AlphaZero against Stockfish), AI nevertheless thanks to more exact estimations of "the relative positions" benefits against is deeper-reboric variant in the form of Stockfish. At some forums specify that played pier Stockfish not at the full capacity since "64 threads" - twice it is less from capacity. ( it is a comment) "with 40ms thinking time" is twice more from the minimum the Full power is thinking 5000! And also that Speak and selected a little (1) for such amount of flows. To include on the declared reference capacity, there it is necessary terabyte and that Stockfish could not use debut basis. And without the debut book, bases, they are forced from zero to count a position, unlike AlphaZero. But and AlphaZero studied only 4 hours, and can after all and month and if hallows does not suffice even after that (what I doubt) - that could be fed also in addition "the debut book, bases" but it is clearly finite that here counter AlphaZero in that that it (in few hours) independently reaches results of the best than all mankind with all chess achievements, including the best programs written to it. Well and so to convergence, AlphaZero and the previous AlphaGo Zero used a single machine with 4 TPUs. And at training, Training proceeded for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters, using 5,000 first-generation TPUs (15) to generate self-play games and 64 second-generation TPUs to train the neural networks.