The Human-in-the-Loop is Tired

Summers argues that LLM-assisted programming is both useful and destabilizing, and that ignoring the destabilizing effect leads to burnout. She writes from inside the team building Pydantic AI, so the piece carries the weight of someone who ships LLM tooling while living with its consequences.

The core observation is a shift in where the cognitive load lands. Hand-coding provided frequent small dopamine hits — solving problems, understanding logic, seeing code compile. LLM-assisted work replaces those with the fatigue of supervision: holding intent in your head while reviewing "mostly-correct output" that suffers from errors of coherence across complex changes. Summers calls this "the human reward function problem" — the satisfying part shrank while the exhausting part grew.

She draws a parallel to the 2009 shift from fixed-width to responsive web design, when designers experienced existential loss of control. Those who thrived reframed their skills around new constraints. Summers argues the pattern holds now, though the stakes and pace are higher — the current shift is measured in months, not years.

What survives, in her view, are taste, nuance, mature architectural opinions, and contrarian calls born of genuine expertise. Teams succeed most when guiding LLMs in domains they deeply understand; in shallower areas, outputs become "impressionistic" — plausible but less correct. Emerging practices include running pre-mortems with fresh LLM sessions and distilling expertise from past code reviews into guidance files.

The piece closes with a reframe: the bottleneck was never the code, but human attention, engineering judgment, and the ability to hold a coherent system vision. Now that code-writing is automated, those capacities are revealed as the scarce resource — and the humans doing this work are tired.