Ending AI Slop: How to Train Models for 'Taste'
AI Engineer · youtube · 2026-08-01
Thais Castello Branco, founder of Taste Labs, discusses how to elevate AI's "taste" in subjective design domains through data construction and reinforcement learning environments, aiming to end generic AI slop.
Core Concepts:
- Decomposition: Breaking down subjective creative work (like brand design) into specific elements that can be individually graded, making the otherwise vague concept of "taste" measurable.
- Combating Collapse to the Mean: Models optimizing for the most likely output often drift toward mediocrity, killing the creativity design requires. The solution is building data and rewards that incentivize breaking from the obvious.
- Building Preference Data: Turning expert judgment into high-signal structured preference data, requiring human reviewers to tie feedback to specific design choices to ensure rich, low-noise preference signals.
She argues that as subjective domains become measurable, taste becomes a dimension that can actually be trained into models.
More from AGI Musings
- Will Automated AI Researchers Become Reality at OpenAI or Anthropic Within Six Months? — firstadopter · 2026-08-01
- Musk Calls Human Language 'Lossy Compression', Predicts Neuralink Will Replace Speech — r0ck3t23 · 2026-08-01
- Netflix CPTO: 'Systems Thinking' is the Most Valuable Skill in the AI Era — lennysan · 2026-08-01
- We Discovered an Industrial Recipe for Intelligence Before Its Science — signulll · 2026-08-01
- Is the Technical Co-Founder Dead? Designer Ships 5 Tools Solo Using AI Agents — npew · 2026-08-01
- Opinion: AI Agents Should Have Their Own Internet — sull · 2026-08-01