Podcast
Why AI Has No Taste and How to Fix It (W/ Thais Castello Branco) | E2319
This Week in Startups
- Why Frontier Models Lack Taste
- Large language models regress to the mean on subjective tasks because training optimizes for the most likely answer, not distinctive aesthetics.
- Thais Castello Branco explains models excel at math/coding but fail at unique design because great work is often out-of-distribution. (Time 0:06:15)
- Regression To The Mean Explains AI Slop
- Regression to the mean makes average outputs ubiquitous so LLMs produce bland, similar designs across distinct prompts.
- Jason Calacanis and Thais note that the training signal rewards likeliest outputs, not creative divergence. (Time 0:07:00)
- Benchmark Models For Aesthetics Before Fixing Them
- Do benchmark and evaluate foundation models for aesthetics and 2D reasoning to find where they break and steer them toward higher-quality outputs.
- Thais says Taste Labs works with frontier labs to measure capability gaps before fixing style issues. (Time 0:09:47)
- Use Paid Tastemakers To Teach Model Taste
- Hire or pay human ‘Tastemakers’ to rate, critique, and curate model outputs so AI learns nuanced judgments and style variety.
- Thais describes Taste Labs’ ~1,000 paid tastemakers who provide critiques, examples, and curated portfolios. (Time 0:15:26)
- Fix Quality First Then Personalize Taste
- Commodifying taste risks homogenization, but first priority is raising the baseline quality (craft and polish) before tackling personalization.
- Lon and Thais agree personalization should follow quality improvements to avoid one-size-fits-all taste. (Time 0:17:10)
- Jason’s Personal Cool Hunter Prompt
- Jason built a personal ‘cool hunter’ prompt that aggregates trusted sources like Monocle to surface boutique hotels and design-forward spots.
- He uses it weekly to get curated trend reports and find artisan hotels and restaurants when traveling. (Time 0:22:50)
- AI Hedge Fund Liquidation And Anthropic Bet
- Leopold Aschenbrenner’s Situational Awareness fund ran heavy leverage and sold public positions to cover margin calls, keeping private Anthropic holdings.
- Jason and Lon discuss the 25-year-old’s 4x leverage, liquidation to Citadel, and survival of private stakes. (Time 0:37:40)
- Google Earth AI Enables Visual Misinformation
- Embedding generative AI directly into Google Earth lets anyone create photorealistic alterations, instantly enabling high-impact visual misinformation.
- Lon and Jason highlight examples: fake refugee camps on borders and fabricated bomb craters near hospitals. (Time 0:50:56)
- Platforms Start Cracking Down On AI Slop
- Platforms are pushing back on ‘AI slop’ by adding detection and moderation tools to prevent low-effort, AI-generated content from filling feeds.
- Lon cites LinkedIn’s report-AI-slop button and Substack’s anti-AI measures to preserve quality. (Time 0:54:26)
- Self Driving Commoditization Appears At Campus Scale
- Self-driving tech is becoming commoditized: students and startups build camera-only autonomous golf carts and small vehicles, lowering barrier to entry.
- Jason points to Ethan Goodhart’s Wind golf cart demo with notable beta riders like Sam Altman. (Time 1:03:44)
- Robots Accelerate Utility Scale Solar Installation
- Robotics and automation can scale solar installation: robots place panels and humans finish bolts, enabling rapid utility-scale deployment.
- Jason highlights Gritt Robotics’ system that unloads, transports, and positions panels to speed large solar field builds. (Time 1:09:48)