Why AI Lacks Taste and How to Fix It
This Week in Startupsgo watch the original →
the gist
AI models regress to the mean because they are trained on average data; Taste Labs is building infrastructure and human-in-the-loop systems to inject aesthetic judgment and unique 'out-of-distribution' quality into AI outputs.
The Problem of Aesthetic Regression
Large language models are highly capable in objective domains like mathematics and coding, yet they consistently fail at subjective tasks like design, writing, and cultural critique. Thais Castello Branco argues this is a direct result of how models are trained: they are optimized to find the most probable, 'correct' answer, which inherently leads to regression to the mean. Because the training data is composed of the average of the internet, the output is predictably bland, derivative, and lacking in unique perspective.
The Role of Taste and Curation
Taste is defined as a rare skill developed through deep, domain-specific exposure and the ability to make high-stakes judgments when no clear 'right' answer exists. The decline of the professional curator class—critics who dedicated their lives to studying specific fields—has left a void. While social media and aggregate review sites have replaced these figures, they haven't replaced the function of high-quality, informed critique. Castello Branco suggests that as AI enables the mass production of content, the 'peak of the peak'—truly unique, high-taste work—becomes exponentially more valuable.
Bridging the Gap with Taste Labs
Taste Labs operates at two levels to solve this. First, they work with frontier labs to benchmark models, identifying where they break and how to steer them toward 'out-of-distribution' quality. Second, they provide tools and APIs for the application layer, helping non-expert users generate better results by providing the necessary context and aesthetic guardrails before generation begins. This involves a 'TasteMaker' community—a network of human curators who provide the preference data required to teach models what 'good' actually looks like in subjective contexts.
The Future of AI-Assisted Creativity
There is a fundamental tension between the democratization of creation and the maintenance of quality. As AI tools become more prevalent, the challenge shifts from 'how do we generate content' to 'how do we ensure the content has a distinct point of view.' Castello Branco emphasizes that we haven't fully solved the personalization problem, where different users should receive different, high-quality outputs based on their specific intent, rather than the same generic 'slop' currently produced by default.