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

Episode Summary

Executive Summary: The episode centers on AI’s current weakness in subjective domains like design, writing, and humor, and argues that “taste” will become more valuable as AI mass-produces average content. Guest Tais Castillo of Taste Labs explains how her company uses expert tastemakers, critique, curation, and preference data to improve AI outputs. The show also covers Leopold Aschenbrenner’s leveraged fund unwind, Google Earth’s risky AI image tools, anti-slop product changes by LinkedIn/Substack, self-driving and robotics commercialization, and large-scale energy/solar optimism.

Main Topics: AI lacks taste in subjective domains (Priority: 5/5): The guest and hosts argue that models are strong at objective tasks like coding and math but still produce generic, average, or “slop” outputs in design, tweets, and creative work because training favors the most likely answer rather than distinctive judgment. Taste as a scarce and rising asset (Priority: 5/5): As AI increases content volume, the hosts argue that true curators, critics, and tastemakers become more valuable, not less, because peak quality stands out more in a world flooded with mediocre output. Taste Labs’ approach to training and productizing preference data (Priority: 5/5): Tais Castillo explains that Taste Labs works with frontier labs and application-layer companies using experts to rate, critique, curate, and create ideal examples, building tools, APIs, and design indexes to improve model behavior and personalization. Leopold Aschenbrenner’s hedge fund unwind (Priority: 4/5): The show covers the AI-themed fund run by 25-year-old Leopold Aschenbrenner, its highly leveraged positions, margin pressure, sale of public holdings to Citadel, and continuation of private bets like Anthropic stock. AI slop, misinformation, and product governance (Priority: 4/5): The hosts react to Google Earth’s Nano Banana integration, LinkedIn’s removal of its AI-writing prompt, and Substack’s anti-AI tools, arguing platforms need stronger friction against low-quality or deceptive AI-generated content. Autonomy, labor disruption, and scaling robots (Priority: 4/5): They discuss self-driving golf carts, Waymo-style commercialization, and drone/robot systems, emphasizing that autonomy is getting commoditized and will pressure drivers and delivery workers, prompting debates over licensing and social compensation. Energy and infrastructure optimism (Priority: 3/5): The episode ends with optimism about solar, batteries, canal-top panels, and automation in utility-scale deployment, suggesting that energy abundance could help resolve broader economic and environmental tensions.

Key Arguments: LLMs are optimized for the average likely answer, which works in coding and math but produces bland creative outputs. Taste requires exposure, repetition, critique, and domain-specific judgment; not everyone can or should develop it everywhere. AI will not eliminate tastemakers; it will increase the value of the best curators because mass production makes peak quality more precious. Improving subjective AI outputs requires both model-side benchmarking and application-layer systems that feed richer context and preference signals. Preference data is highly valuable and can be collected through ratings, critiques, curated examples, and idealized exemplars rather than simple labels alone. Platforms should create stronger anti-slop mechanisms, because AI-generated content can degrade feeds, trust, and user experience. Leverage, not just market volatility, caused Leopold Aschenbrenner’s fund to unwind; the story is more about risk management than a fundamental fraud analogy. Self-driving and robotics will scale quickly, but their social adoption will be constrained by labor backlash and the need for policy smoothing mechanisms. Large-scale solar and automation could drastically lower costs and reduce conflict over energy and water if deployed at scale.

Data Points: Age of Leopold Aschenbrenner: 25 - The hedge fund manager discussed in the news segment Fund assets under management: over $45 billion - Reported AUM for Aschenbrenner’s fund by July 2026 Fund return through June 30: 439% net return - Performance cited before the unwind Initial public market sale buyer: Citadel - The public stock portfolio was sold to Ken Griffin’s firm Private investment exposure retained: $5 billion in Anthropic stock - Financial Times-reported remaining major private position Seed round: $18.5 million - Taste Labs funding round Community size: about 1,000 tastemakers - Taste Labs’ expert network across design/media styles First six hours free: 6 hours - YSecurity ad offer for startups One-bedroom rent in San Francisco: $3,500 to $4,600 per month - Host discussion about housing costs and hiring in SF Population entering Ceuta: 60,000 - Spanish government figure for migration pressure in the enclave Ceuta resident population: 85,000 - Context for scale of migration surge Reported deaths at the border: about 18 - Deaths attributed to drowning or stampede during crossing attempts Solar share of California energy: 51% - Host cited recent renewable milestone Water evaporation reduction from canal solar: 70% - Claim about solar panels covering irrigation canals AI preference threshold mentioned: over 50% / 90% confidence - Hosts proposed ranking or gray-out threshold for suspected AI content

Pivotal Quotes: "The people who build these things have no taste, let's be honest. They don't have taste." — Jason Calacanis: Opening riff on why AI-generated designs and writing often feel generic "The more you have of this mass production of things, the more I think the peak of the peak becomes valuable." — Tais Castillo: Guest explanation of why tastemakers gain importance as AI output volume rises "We came closer to permanent capital impairment than is acceptable to us." — Leopold Aschenbrenner: Excerpt from the LP letter describing the fund’s leveraged unwind

Implications: Listeners should expect a push toward taste-aware AI systems, stronger anti-slop moderation, and rising demand for human curators. The episode also signals continued volatility in AI investing and faster automation pressure on labor, even as infrastructure and energy innovations scale.