“While AI model loss improves linearly and predictably with scale, specific capabilities like arithmetic can emerge discontinuously, jumping from inability to reliability.”