Toward Computational Taste: LLMs, Aesthetics and Judgment (patron.fund)

The essay argues we’re entering a new regime where large language models and generative systems don’t just reflect human taste—they learn, optimize, and eventually shape it. Taste, once mediated by curators and social networks, is now computable: datasets (Pick-a-Pic, ImageReward, HPSv3) and methods (TAPO, G‑Eval) let models approximate human judgments of style and appeal, while personalization approaches expose the limitations of single‑reward RLHF. The piece frames this shift as culturally and economically consequential: models that understand “taste” enable far more context-aware personalization, novel product categories, and new forms of cultural influence. Technically, the writeup highlights low‑rank reward modeling (LoRe) and collaborative ranking as practical ways to represent individual preferences as combinations of shared basis functions, enabling few‑shot personalization without full fine‑tuning. From taste‑conditioned generation and “Taste-as-a-Service” APIs to a structured “Taste Graph” that maps cultural capital, the implications span recommender systems, creative copilots, and social platforms organized by style rather than social graph. The author also flags real risks: freezing users into past preferences, commodifying cultural capital, and concentrating power in whoever builds the taste models. The takeaway for AI/ML practitioners: modeling taste is a feasible, high‑impact frontier that demands new technical tools, careful personalization design, and ethical safeguards.