Taste
Towards a Dynamic Model of Taste proposes that taste should be treated with inferential seriousness. The prompt is Andrej Karpathy’s recent observation that, as agentic AI systems do more delegated work, humans remain responsible for aesthetics, judgement, taste and oversight. Machine learning advances fastest where outputs can be verified: code compiles, tests pass, games are won, and proofs check. Taste is harder because its reward signal is plural, delayed, contested and historically unstable; but it has its metrics.
The paper models taste as a generative field rather than a scalar value or an imposed ideological construction. A work is shown to have a seven-axis profile: aliveness, tradition, reshaping, bearing, crediting, witness and generation. But that profile becomes visible only as it moves through readers, critics, teachers, markets, institutions, BookTok communities, later artists and time. Different communities apply different weights; sales, reviews, prizes, teaching, imitation and adaptation all measure something, but no single measure exhausts value. Value is both resilient and fugitive; it is culturally variable, but not merely elusive.
The model therefore distinguishes preference from discrimination and offers ways to think about how meaning changes over time, and about who controls the bank of semantic value. A reader may like a work; a ‘trained’ judge may also perceive form, cliché, tradition, pressure, falseness, necessity and generative power. In the age of abundant machine production, the scarce act is not first output but judgement: what is selected, rejected, revised, credited and released into the world. The provocation is simple: computer science may now need poetry, because literary culture has spent two and a half millennia studying judgement where no clean reward signal exists, and how value changes across communities, institutions, markets and time.