"Human performance" on Atari games isn't how well a human can play with their own 10 fingers. It's how well the abstract mental models of the game inferred by the average human can perform when coded up into a simple program. That means perfect scores for most games.
And the data-efficiency of that process is staggering -- a human only needs a handful of games to come up with mental abstractions enabling it to programmatically solve most of these games at 100%. That's what the bar is. That's what we should aim for.
Don't let them tell you that deep learning has achieved "superhuman" performance at any of these games -- any random programmer can come up with a better solution program for a given game in an afternoon. That's what human-level means.
