Research reading

Is taste a moat?

SOURCE AUTHOR / MICHAEL MALYUK

Notes on the source essay supplied with this study. These are report summaries, not the complete original articles.

When teams can access the same models, knowing what good looks like becomes a practical question. Human judgment matters, but a preference label on its own does not explain that judgment.

The source essay asks whether taste can become a system that continues to learn. Code has compilers and tests. Creative decisions need other instruments: examples, standards, comparisons, and people who can explain what they see.

Consensus can establish a quality floor. It can also average away the differences that make an output useful for a particular audience. A preference study should retain those differences instead of treating the majority choice as a universal standard.

This report applies that idea by keeping the brief and the model’s stated decisions beside the work. The global preference ranking is one view of the results. It should be read alongside the variation between briefs and, when available, the individual annotations.