AI's Quest for Taste: Bridging Data and Design

The evolution of AI-generated design is reaching new heights, with experts like Ben Blumenrose observing significant improvements in the quality of outputs from AI models. As the co-founder and managing partner at Designer Fund, Blumenrose has witnessed firsthand how AI tools have advanced, noting that they are now producing designs that are five times better than just a year ago. The gap between human and AI-generated designs is closing rapidly, and Blumenrose predicts that soon, even seasoned professionals will struggle to differentiate between the two.

As AI technology progresses, the industry is now shifting its focus toward a more complex challenge: instilling a sense of 'taste' in these models. Historically, AI has been evaluated on its ability to avoid errors—like ensuring generated images have the right number of fingers or that essays maintain coherent metaphors. However, as baseline performance becomes the norm, the quest for creative appeal has emerged as a new frontier.

Startups and major tech firms alike are working to encode taste into their AI systems, but this endeavor is fraught with challenges. The subjective nature of taste means that opinions differ widely based on personal experiences and cultural backgrounds. While developing a model’s taste involves extensive data collection and analysis, it also requires a nuanced understanding of what constitutes appealing design. This balancing act is proving to be both a technical and an artistic challenge.

To cultivate AI taste, companies are gathering domain-specific data rather than relying solely on generalized knowledge. Figma, for example, recognizes that off-the-shelf models often lack the sophistication to evaluate design quality reliably. Sumithra Bhakthavatsalam, Figma’s AI Research Lead, explains that the company is focused on improving output speed and quality through a process called post-training, which involves collecting curated examples of effective designs from its user base.

Meanwhile, Krea, a platform catering to creative professionals, is dedicated to developing high-quality visual outputs. Co-founder Diego Rodriguez emphasizes the importance of gathering specific references in film and architecture to ensure their AI models are equipped to meet professional standards. A recent collaboration with architectural firm Henning Larsen has further refined Krea’s offerings by integrating expert feedback into their model training.

The challenge of teaching AI discernment goes beyond technical improvements. Edwin Chen, CEO of Surge, suggests that instilling taste in AI is a fundamentally humanistic endeavor. It’s not simply about providing facts; it’s about teaching models values and preferences. This is where reinforcement learning with human feedback (RLHF) comes into play, allowing AI to refine its outputs based on assessments from experienced professionals. This feedback loop is crucial for guiding AI toward more tasteful designs.

As AI continues to evolve, the industry is at a pivotal crossroads, trying to define and measure taste in a way that resonates across various creative fields. The collaboration between human experts and AI technology may redefine how we approach design, making it an exciting time for both creators and technologists.

Source: www.fastcompany.com