何韵中 / He Yunzhong

Overview

何韵中 is presented in the source as a Scale AI researcher working on model post-training and evaluation.

Current Profile

In the episode, He explains the AI-data business as a changing production system rather than a fixed labeling market. His account connects expert-authored rubrics, model-appropriate task difficulty, software environments, verifiers, successful trajectories, and training recipes. He stresses that code and mathematics became early reinforcement-learning domains because outcomes are relatively checkable, while medicine, finance, creative work, and long-horizon decisions require richer combinations of rules, model judgment, and human review.

His commercial argument is that laboratories can synthesize data internally yet still buy from suppliers that acquire private workflows, negotiate software and data rights, recruit experts, and discover new domains. These are source-reported industry judgments, not a complete description of Scale AI’s products or economics.

Key Characteristics

• Post-training and evaluation researcher speaking from the Scale AI context.
• Treats rubrics and agent environments as complementary layers rather than competing data products.
• Emphasizes task-difficulty matching, verifier strength, and successful trajectories in model training.
• Frames procurement, licensing, expert access, and continual research as supplier advantages.
• Distinguishes meaningful benchmark improvement from leakage, narrow specialization, and reward hacking.

Evidence

• Role and scope - E253｜谁在给大模型出题、卖题、判卷？聊聊AI数据行业的野蛮生长 introduces He as working on post-training and evaluation research at Scale AI.
• Training-system view - E253｜谁在给大模型出题、卖题、判卷？聊聊AI数据行业的野蛮生长 attributes to the discussion the integration of tasks, tools, sandboxes, rubrics, programmatic checks, and trajectories.
• Data-market view - E253｜谁在给大模型出题、卖题、判卷？聊聊AI数据行业的野蛮生长 records arguments about third-party procurement, exclusive data, vertical expertise, and research-led product turnover.

Qualifications

The source provides no detailed biography, publication record, employment dates, or quantitative evidence for the market forecasts attributed to the conversation. His role and claims are preserved at the scope supplied by the episode.

What Changed

• Added He as a source for the wiki’s post-training, verifier, and AI-data-market branches.

Relationships

• Scale AI - employer and research context stated by the source.
• 孙一游 - co-guest discussing benchmarks, environments, and data incentives.
• Agent’s Last Exam - evaluation project used as a cross-domain case.
• Expert Rubric Verification / 专家评分标准验证 - method he helps distinguish from the execution environment.
• Vertical AI Data Procurement / 垂直 AI 数据采购 - commercial bottleneck emphasized in the discussion.
• Environment-Based Agent Benchmarks - broader technical frame for tasks, tools, sandboxes, and verifiers.