When he got back from Pittsburgh, there was nothing left to do. Outside, Toronto was shedding its winter coat and gearing up for an abnormally hot summer, its elms and oaks and maples silently sprouting the green silk of life. In Ray’s minimally furnished apartment, light flooded the living room windows and illuminated his computer screens, turning them into Petri dishes of dust and smudged fingerprints. Even though the bot itself had not achieved perfection, and had to use small approximations of its own, there was nothing Ray could do to reach its level; it just couldn’t be done. Deep inside, machine learning systems were not individuals, Cardanus was not one brain: the evolution of artificial intelligence had lately converged toward multiagent systems, nets of individual nodes working simultaneously and collaborating in a kind of swarm intelligence. Tasks that could proceed separately, with no mutual dependency or even communication, were called “embarrassingly parallel.” The difference between an agent’s performance results and optimal performance was called “regret.” Cardanus, with its myriad interconnected cells working by trial and error, free of the shackles of corporeity, had brought to the table a game-changing regret minimization technique that Ray, alone in an apartment filling up with summer, had no way to possibly emulate. It wasn’t a fair fight. And it was over.