This episode explores Thais Castello Branco’s journey founding Taste Labs, a company aiming to solve AI’s lack of “taste” by training models to produce nuanced, context-aware outputs rather than generic “slop,” while addressing the philosophical and technical challenges of subjective quality in machine-generated content.
Why she went solo
Thais decided to start Taste Labs alone after realizing the idea was too urgent to wait for a co-founder, despite initially considering partnership options.
She had prior startup experience, including co-founding a company in Brazil and joining Exa’s early team, which shaped her founder mindset.
The transition from Exa to Taste Labs felt natural, with rapid fundraising and hiring allowing her to move forward without a co-founder.
She emphasizes that solo founding requires high pain tolerance and emotional resilience, but offers clarity of vision and conviction that she believes is essential for mission-driven work.
What “taste” actually is (and why it’s not personalization)
Taste is defined as a judgment of quality in subjective domains, requiring shared validity among experts rather than individual preferences.
It is a skill developed through years of practice, pattern recognition, and deliberate curation of what to include or exclude.
Unlike personalization (tailored to one person), taste involves understanding what “great” looks like in a domain and recognizing when something meets that standard.
The ability to develop taste is rare because it demands sustained effort and exposure to diverse examples, not just natural talent.
The real cause of slop: volume and lack of context
AI-generated “slop” stems from high-volume, low-context outputs that feel thoughtless and repetitive across different use cases.
When millions use the same tools, quirks become homogenized rather than unique signatures, creating a monoculture effect.
The gap between AI output and human expectations is exacerbated by vague prompts and users lacking the vocabulary to articulate their intent.
Taste Labs aims to bridge this gap by teaching models to understand situational appropriateness and diverse stylistic preferences.
How machines learn taste differently than humans
Humans develop taste through continual, organic learning—imitating forms, experimenting, and refining over time (e.g., Picasso’s evolving style).
Machines currently learn in batches (pre-training, post-training) rather than through ongoing, contextual adaptation.
Subjective domains like design or writing require nuance and situational awareness that machines struggle to replicate without explicit training.
The goal is to create models that can recognize patterns of quality while adapting to specific contexts, not just averaging preferences.
Inside the ~800-person tastemaker community
Taste Labs collaborates with 800 specialists across visual design, including websites, slides, and generative art, to define quality standards.
These “tastemakers” help create rubrics for grading outputs, distinguishing between bad, okay, and great work.
The community emphasizes diversity in backgrounds and styles to ensure training data reflects varied perspectives, not just dominant trends.
The process involves both creators (who make great work) and critics (who explain why it’s great), recognizing these as distinct but complementary skills.
The Lindy effect and what lasts
Thais believes that high-quality, intentional work will stand out more in an AI-saturated world, as “okay” becomes easy to produce.
She cites examples like Pike Place Market and long-standing brands (e.g., Fishwife, Liquid Death) as embodying authenticity and craftsmanship.
The Lindy effect—longevity predicting future survival—suggests that enduring creations often have deeper care and intentionality.
While AI may increase the volume of ephemeral content, she argues that truly exceptional work will still emerge and persist if it reflects genuine creativity and purpose.
Why hasn’t taste been solved?
The problem remains unsolved due to its ambiguity and technical complexity, with many open questions about how to train models on subjective quality.
Thais argues that progress, not perfection, is the goal, and that avoiding the problem risks normalizing low-quality outputs.
She emphasizes that taste Labs is not dictating a single standard but enabling models to reflect the diversity of human judgment and context.
