The remarkably human task of giving AI ‘good enough’ taste

Ben Blumenrose, co-founder and managing partner of Designer Fund, a venture capital firm specializing in startups, has observed the rapid advancement of AI-generated visual design. Despite his ability to distinguish between human-made and AI-made websites, Blumenrose believes that the difference between AI-produced designs and professional human-made designs will become increasingly difficult to discern within the next six months to a year.

Framer, a no-code website design platform and one of Designer Fund's investments, represents this shift.

The AI industry has been focused on improving the creative output of models through evaluation metrics, rubrics, and automated tests. However, as "correctness" becomes the baseline, the challenge now lies in creating work that impresses experts like Blumenrose. This involves developing an approximation of taste or creativity.

Developing taste in AI systems is a complex task as opinions on what constitutes "good" can vary greatly depending on personal and professional experiences. Just like in traditional creative fields, learning the rules and understanding what one likes is crucial. Similarly, teaching AI taste involves gathering relevant information and making discerning choices.

To create AI with a refined taste for specific creative fields, companies are focusing on collecting domain-specific expert data. Companies like Figma and Krea are curating high-quality examples of "great" and "good" designs for their designers and users. By post-training these models with curated data, they aim to improve the output quality, variety, and overall understanding of the specific domains they serve.