saved
Building software factories (with no slop)
gist
David argues that AI slop is a verification bottleneck, not a model-quality failure: generation outruns human reading, so review becomes a formality. The way out is to stop treating first-party code as something a human must read. Structure the system into domain-specific pieces with seams, sensors, and living artifacts—invariants, traces, attack surface, and a ranked ledger of silent decisions—so engineers inspect intent and behavior while the implementation can become unreadable. A software factory is that production line, not a coding agent.
ideas
- Slop is a verification failure. Models can write code; unbounded generation plus human-speed reading makes LGTM a formality.
- Human reading cannot stay the filter. Matching generation to reading speed throws away the advantage; embracing slop accepts the degradation.
- Treat the codebase as a black box. Domain-specific pieces with defined I/O, sensors, and seams make behavior inspectable without reading implementations.
- Review decisions, not lines. A separate auditor returns a ranked ledger of silent choices; the implementer must not audit or patch itself.
- Slice until pieces fail independently. Scout fog of war, carve independently verifiable territories, and re-slice when the map hides more map.
quotes
“Slop is what happens when generation is unbounded and verification is bottlenecked on a human reading the output.”
“The interface becomes the review surface.”
“I read the thirty. I push back on four.”
“That's what a software factory is, to me.”