Human Taste Can Be Gamed: The Challenges of Verifiable Preferences
zacharynado · x · 2026-08-25
Torchcompiled discussed how human taste can be gamed, leading to empty supernormal stimuli. The argument counters the verifiability of taste, noting preferences are continuous/multi-axis, often intransitive, change with discovery, are socially manipulated, and nonstationary. This touches on core challenges in RLHF and recommendation systems.
Related event: Debate: Human Taste Can Be Gamed, Not Verified(2 posts)→
More from AGI Musings
- Models lack independent research capabilities; RSI predictions seem overly optimistic — BlancheMinerva · 2026-08-25
- Anthropic CEO: 'No Wall' for AI Synthesizing Cross-Domain Knowledge — rohanpaul_ai · 2026-08-25
- Computational thinking becomes essential as agents handle execution — every · 2026-08-25
- Opinion: Math Must Shift to Macro View to Compete with Machines — yeastsplainer · 2026-08-25
- How does philosophy background influence views on progress? — PierceLilholt · 2026-08-25
- Economist observes AI's tendency to invent jargon, sacrificing readability — paulnovosad · 2026-08-25