Why was Go harder for computers than chess?
Because the techniques that solved chess do not scale to Go, and the position is far harder to evaluate — so Go required a genuinely different approach rather than more computing power.
The branching factor. Chess has roughly 35 legal moves in a typical position. Go, played on a 19×19 board, has around 250. Because search trees grow exponentially with branching factor, looking a few moves ahead in Go requires astronomically more computation than the same depth in chess. Brute-force search simply cannot reach useful depth.
The number of possible positions in Go is frequently described as exceeding the number of atoms in the observable universe — a comparison that is unhelpfully large but conveys the point: exhaustive approaches are not merely impractical, they are impossible in principle.
The harder problem: evaluation. In chess you can score a position reasonably well with simple heuristics — count material, assess king safety, and you have a usable estimate. In Go, no comparable shortcut exists. All stones are identical, so there is no material to count. Whether a group of stones is strong or dead, whether territory is secure, and whether a shape is efficient are global, subtle judgements that resist being written as rules. Strong human players describe evaluation in terms of intuition and shape recognition rather than calculation.
What eventually worked. Two developments combined:
Monte Carlo Tree Search. Rather than evaluating positions with rules, play out many random or semi-random continuations to the end and use the proportion of wins as an estimate. This sidesteps the evaluation problem entirely, and produced the first meaningfully strong Go programs in the mid-2000s.
Deep neural networks. AlphaGo combined MCTS with networks trained first on human games and then through self-play: a policy network proposing plausible moves, dramatically narrowing the search, and a value network estimating who is winning. It defeated Lee Sedol in 2016, a decade earlier than most expected.
AlphaGo Zero then learned entirely from self-play with no human games, and surpassed the earlier version — suggesting human knowledge had been a constraint as well as a starting point.