Human Taste as AI Training Signal / 人的品味作为AI训练信号
Human taste as AI training signal is the source’s claim that people still matter because they decide which problems are important, what direction is better, and what counts as a high-quality answer. In 174. 我们还能给算法当多久的品味老师？｜对谈亚马逊AGI查晟, 查晟 / Cha Sheng says current self-improvement loops still need human taste to define goals, set evaluation standards, and tell AI systems the shorter or more useful path.
The sharper point is that taste is not protected simply because it is human. Once problem choice, standards, examples, reviews, preferences, and reasoning can be written as text or captured as data, they can become training material. The source therefore extends Research Taste and Human Judgment Under AI: taste remains a bottleneck, but it can also be distilled, imitated, averaged, or commoditized by models.
Key Claims
- Taste includes knowing which questions matter, why they matter, and which solution direction is better.
- Human taste remains valuable where problems are novel, ambiguous, sparse in data, or require cross-domain context.
- Common-domain taste can become less scarce when enough human standards are expressed in documents, feedback, reward data, or evaluation rubrics.
- Expert taste may become more valuable as ordinary taste is absorbed, because frontier systems still need new standards and out-of-distribution problem framing.
- Taste becomes operational only through an Agent Harness or workflow that turns judgment into tasks, tools, reviews, and feedback.
Connections
- 查晟 / Cha Sheng and Amazon AGI - source speaker and model-building context.
- Research Taste - adjacent research-method concept.
- Human Judgment Under AI and Human Agency Under AI - broader human responsibility and agency boundary.
- Content Engineering, AI Answer Evaluation, and Data As Education - nearby wiki frames where human standards are turned into model behavior.
- Recursive Self-Improvement and Agent Self-Evolution - self-improvement loops that still need taste and verification.