查晟 / Cha Sheng

Cha Sheng is the Amazon AGI research-team lead interviewed 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 about foundation-model training, open and closed model competition, synthetic data, self-improvement, AI values, enterprise models, and AGI limits. In the source, he is presented as someone who has participated in multiple Amazon large-model training cycles and can speak from inside the basic model-training workflow.

His recurring frame is that current AI remains dependent on people for problem definition, taste, evaluation, and organizational direction, while also making those same human contributions easier to distill into model behavior. That makes him an important source for Human Taste as AI Training Signal / 人的品味作为AI训练信号, Research Taste, Agent Harness, and Human Agency Under AI.

Key Points

• He treats Open Source AI Models as strategically important but structurally weaker than closed consumer products at capturing direct user data.
• He frames synthetic data through information quality, filtering, and evaluation rather than through a simple synthetic-versus-human binary.
• He argues that model values emerge from data, rewards, and training choices, connecting his comments to Model Value Embedding / 模型价值观嵌入 and AI Alignment Governance.
• He sees future AGI as requiring fast learning, Continual Learning, and reduced forgetting, not only larger static language models.

Connections

• Amazon AGI and Amazon - team and company context.
• 起朱楼宴宾客 / Qizhulou Yan Binke and 大卫翁 / David Weng - interview setting.
• Human Taste as AI Training Signal / 人的品味作为AI训练信号, AI Knowledge Collapse / AI知识塌缩, Cognitive Debt / 认知负债, and Sovereign AI Models / 主权AI模型 - concepts introduced or sharpened by the source.