Ending AI Slop Through Domain Decomposition

The Breakthrough

Subjective domains like design and creative writing suffer from 'collapse to the mean' because models optimize for the most likely output. The breakthrough is to treat these domains as routing problems, decomposing fuzzy tasks into verifiable sub-elements (such as typography, color, or spacing) to create ground-truth signals that prevent the model from drifting toward average, low-quality outputs.

What Actually Worked

• Decompose subjective assets into codified, verifiable components to create an RL environment where the model is graded against specific design constraints rather than a vague 'quality' metric.
• Implement a routing strategy to determine which problems are best solved via RL environments (verifiable tasks) versus those requiring human-preference data (subjective, contextual tasks).
• Curate preference data by pairing human expert commentary with specific code components or visual elements, ensuring the model learns the relationship between a design choice and the underlying implementation.
• Use expert consensus as a QA mechanism: treat disagreement on objective fundamentals (e.g., alignment) as a data flaw, but treat disagreement on stylistic choices as valuable signal that captures the pluralistic nature of human taste.
• Prioritize high-signal, expert-curated data over large volumes of noisy data, as subjective domains require intentionality to avoid the statistical average.

Context

AI models excel at verifiable tasks like coding and math because those domains have clear ground truths and decompose easily. In contrast, design and writing are subjective, contextual, and change over time. When models are trained on generic preference data, they default to the most likely outcome, resulting in 'slop.' By breaking these domains down into measurable, on-brand components, engineers can force models to move away from the average and toward intentional, high-quality creative outputs.