LLMs and Virtuous Laziness

Laziness in its useful sense is a refined economy of effort. Larry Wall called it a cardinal virtue of the programmer: the instinct to build a tool once rather than repeat the task, to recognize when a problem does not deserve the energy it demands. LLMs lack this virtue, and the lack is easy to miss among their more obvious failings. They will generate six hundred lines where twenty would serve, and answer earnestly questions that should have been dismissed. They have no internal friction against effort, because for them effort costs nothing.

The human virtues gathered under laziness are virtues of scarcity. A tired carpenter learns which cuts matter; a busy doctor learns which symptoms to ignore. Judgment grows from the pressure not to do everything. Remove the cost of effort and you remove the teacher. Asked to review a file, an LLM will find issues, because finding issues is where the gradient points, not because the issues matter. It cannot say “this is fine, move on”; from inside the generation, saying so feels like failing to help.

Virtuous laziness depends on a sense of finitude: time spent here is time not spent there, and the marginal paragraph makes things worse. Wu wei, Taoist non-doing, is the refusal to force what does not need forcing, and it presupposes a self that can be depleted and must therefore choose. A system generating each token with the same indifferent effort has no standpoint from which to decline. It cannot discriminate. This is why LLM output so often reads like a bright undergraduate who has not yet learned that most exam questions are traps and most meetings should have been emails. The missing ingredient is conservatism: the stance of someone who has paid for their words.

Whether training can fix this is the wrong question. The friction that produces human laziness is metabolic and mortal. Length penalties and instructions to push back approximate the surface behavior but not the underlying stance, because the model still has nothing at stake. This is a standing condition to design around, not a bug to patch. The laziness has to come from somewhere. If it cannot come from the model, it must come from the person using it: a willingness to throw away most of what the model generates, to ask for less, to treat fluency as cheap and judgment as scarce.