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

Thais Castello Branco, founder of Taste Labs, argues that AI models optimized to find the most likely correct answer are structurally guaranteed to produce regression-to-the-mean outputs for subjective tasks like design and writing, because great design and great writing tend to be out-of-distribution rather than average. This explains why current AI can solve PhD-level math and cybersecurity problems but cannot write a good tweet or produce a good design. The missing ingredients include variety, fit to prompt, fit to user intent, and personalization, with different users submitting different prompts often receiving outputs that look identical.

Taste Labs is building a two-pronged business to close this gap. One line works directly with frontier labs to benchmark and improve models on foundational capability gaps such as vision and 2D reasoning as well as higher-order qualities like creativity, style, and aesthetics. The other line works with application-layer and agent companies including Lovable, Krista, Codex, Claude, and Figma Design to improve outputs without touching the model layer by feeding the right context before generation begins. About half of the business resembles data labeling work covering design critique, curation of good and bad examples with explanations, and creation of ideal reference outputs compared to LLM-generated ones, while the other half is tools and API infrastructure including a judgment system to verify whether outputs stay on brand.

The company employs roughly one thousand taste makers across different design domains, media types, and styles, paying them approximately fifty to one hundred dollars per hour. A survey of contributors found they are motivated primarily by curiosity and a desire to prevent AI homogenization rather than financial incentives. Taste Labs raised an 18.5 million dollar seed round co-led by CRV and Amplify and is currently focused exclusively on visual design. Castello Branco distinguishes between elements of quality most people agree on, such as craft and polish, and elements people disagree on, such as style and intended audience, and says the company's first priority is lifting baseline quality before tackling personalization.

Castello Branco argues that taste makers and creative directors will become more valuable as AI enables mass production of outputs, not less, and that the company is infinitely far from replicating the power of true creative direction with AI. Calacanis raises the concern that commodifying taste through AI removes the rarity that makes taste valuable. Castello Branco acknowledges the half-life of cool has always existed independently of AI or social media, but says social media accelerated that cycle and AI has the potential to do the same, though it has not yet because AI taste output remains too poor. She argues that if AI pushes diversity of aesthetic directions rather than compressing toward the mean it could help rather than harm the problem, and that the goal is explicitly not to make AI output the same recommendations to everyone.

Leopold Aschenbrenner, a 25-year-old former OpenAI employee, launched an AI-focused hedge fund in 2024 backed by the Collison Brothers, Daniel Gross, and Nat Friedman, holding positions in SK Hynix, Micron, Sandisk, CoreWeave, and Nebius Group. The Wall Street Journal reported the fund reached 20 billion dollars in assets, and by July assets under management had grown to over 45 billion dollars with a reported 439 percent net return through June 30th. Aschenbrenner was running approximately 4x leverage, was margin called by Goldman Sachs and Bank of America, and sold the public portion of his portfolio to Ken Griffin's Citadel to access cash while retaining approximately 5 billion dollars of Anthropic stock. He sent a letter to LPs acknowledging they came closer to permanent capital impairment than acceptable. Calacanis attributed the core problem to the leverage structure rather than inexperience alone and predicted Aschenbrenner will come back stronger given his remaining Anthropic holdings.

LinkedIn removed its AI-assisted post writing button after users were posting AI-generated content at scale and added a report AI slop button. Substack added anti-AI tools and its CEO explicitly said he did not want Substack to become like LinkedIn. Google Alphabet embedded an AI image generator into Google Earth that was quickly used to create fake refugee camp and fake bomb crater images, which Calacanis and Castello Branco attributed to a company-wide impulse to add AI across all products without strategic purpose. The prediction was made that flagging and shadow banning AI-generated content above a detection threshold will become a major trend across social networks.

Ethan Goodhart, a Stanford computer science graduate, built a self-driving golf cart using only cameras and vision models with no LiDAR, with Sam Altman, Andrej Karpathy, and Jensen Huang among early beta testers. Calacanis described Tesla as furthest ahead on self-driving measured by a double nines reliability metric and noted Zoox received federal approval for 2,500 units with no steering wheel, no pedals, and no driver seat. China has stopped issuing new autonomous vehicle permits, which Calacanis attributed to job protection concerns rather than any specific safety incident. He proposed that self-driving companies pay a licensing fee of roughly one thousand dollars per month per vehicle deployed, with proceeds funding retraining or direct compensation for displaced drivers.