The speed at which we managed to progress from no communication to full natural language Diplomacy also surprised the research team. When we started, the idea of an AI agent that could master no-press Diplomacy seemed like a multi-year effort, and the idea of an AI agent that could play full-scale Diplomacy in natural language seemed like science fiction. We thought it might take 10 years to reach this point. -NB

I think the speed that it went from playing no communication games to full communication games was the biggest surprise - not just the natural language but the adaptation of strategy and tactics. I expected it to really struggle to climb out of what it had learned from that style of game, but it did so pretty quickly, which is probably down to the technical expertise of the team. I guess beyond that, the AI plays some approaches that upset the inherited wisdom of the diplomacy playing group. I'm totally revisiting some opening lines for example. In terms of what we can learn - the strategic ideas the emerge seem to be very much aligned with high level human players. Patience, collaboration, improving position rather than brute force tactical tricks... at that level of abstraction it plays very similarly to a good human player. -AG