Why AI has no taste and how to fix it (w/ Thais Castello Branco) | E2319

What It Covers

Thais Castello Branco, founder of Taste Labs, joins This Week in Startups to explain why AI models excel at objective tasks like coding and math but consistently produce mediocre visual design and creative writing. Her company's two-pronged approach combines frontier model training with a community of roughly 1,000 paid domain experts to push AI outputs away from regression to the mean.

Key Questions Answered

• •AI taste failure root cause: Large language models are trained to find the most statistically likely correct answer, which works for math and coding but produces mediocre creative output. Great design and writing are inherently out-of-distribution — they are unusual, not average. The same prompt architecture that makes AI reliable at objective tasks actively prevents it from generating anything distinctive or memorable in subjective creative domains.
• •Two-sided business model for taste data: Taste Labs operates two parallel revenue streams: selling preference datasets and benchmarking services directly to frontier labs like Anthropic and OpenAI, and providing API-layer tools to application companies such as Lovable and Figma that cannot modify underlying models. Founders building AI-native products should consider which layer they can realistically influence before deciding where to invest in quality improvement.
• •Hiring domain experts as "tastemakers": Taste Labs pays approximately 1,000 specialists — senior designers, creative directors, and critics — to critique AI outputs, curate good and bad examples with written explanations, and generate reference-quality work from scratch. Some have transitioned to full-time roles. This structured human-in-the-loop approach produces richer preference signal than binary thumbs-up ratings, which is the dominant data shape most labs currently collect.
• •Personalization as the next quality frontier: Current AI outputs from different users with entirely different prompts often look identical. Thais frames this as a two-stage problem: first lift baseline quality by reducing obvious craft failures, then solve personalization so outputs reflect individual style, audience, and intent. Founders building creative tools should prioritize fixing foundational quality before layering personalization features, since personalization on top of mediocre output compounds the problem.
• •Curators become more valuable, not obsolete: As AI enables mass production of creative content, the scarcity and value of genuine taste judgment increases rather than decreases. Taste Labs survey data from their tastemaker community shows participants are motivated primarily by wanting to solve the slop problem, not financial compensation alone. Platforms and companies that identify and compensate genuine domain experts now will build defensible data moats as preference data becomes the primary differentiator between frontier models.
• •Leopold Aschenbrenner's Situational Awareness fund collapse mechanics: The 25-year-old former OpenAI researcher built a fund from a few hundred million to $45 billion AUM with a 439% net return through June 2025 by taking leveraged long positions in AI infrastructure stocks including SK Hynix, Micron, and CoreWeave. Running approximately 4x leverage meant a 25% market decline triggered full liquidation. The public portfolio was sold to Citadel; roughly $5 billion in Anthropic private shares remains. Leverage at this ratio is the identified failure point.