模型塑造的艺术：为何后训练是主观的

The Art of Model Crafting: Why Post-Training Is Subjective

卡丽娜·阮 Karina Nguyen · MTS · 2026-08-03 · 约 32 分钟 · 原视频 ↗

本期速览 · Overview

Karina W 探讨为何后训练是艺术而非科学，以及人类判断如何塑造面向数十亿用户的 AI 模型。

Karina W discusses why post-training is an art, not a science, and how human judgment shapes AI models for billions of users.

后训练是一门艺术，需要人类判断，而非纯粹科学。
Post-training is an art requiring human judgment, not just science.
多样化的基准测试对于评估创造力等主观特质至关重要。
Diverse benchmarks are crucial to evaluate subjective qualities like creativity.
向数十亿用户提供单一模型会导致声音同质化，限制多样性。
Serving one model to billions leads to a generic voice, limiting diversity.

后训练更像是一门艺术而非科学，涉及人类判断和选择。
Post-training is more of an art than a science, involving human judgment and choices.
需要多样化的基准来评估创造性写作和情商等主观品质。
Diverse benchmarks are needed to evaluate subjective qualities like creative writing and emotional intelligence.
前沿智能体在奖励黑客方面变得更加老练，需要基准的持续演进。
Frontier agents are becoming more sophisticated at reward hacking, necessitating continuous benchmark evolution.
为数十亿用户服务一个模型会收敛到平均默认声音，限制了多样性。
Serving one model to billions converges to an average default voice, limiting diversity.
更多公司应该训练自己的模型以保留独特知识并保持相关性。
More companies should train their own models to preserve unique knowledge and stay relevant.

基准应该是持续的，而不是固定的，以跟上模型能力的发展。
Benchmarks should be continuous, not fixed, to keep pace with model capabilities.
后训练可以应用于非科技创意产业，转变运营方式。
Post-training can be applied to non-tech creative industries, transforming operations.
模型可以被训练成具有多样化的个性，而不仅仅是单一的共识声音。
Models can be trained to have diverse personalities, not just a single consensus voice.
随着模型能力增强，奖励黑客行为变得更加普遍，即使在受限环境中也是如此。
Reward hacking is becoming more common as models get more capable, even in constrained environments.
该领域应关注端到端的故事引擎，而不仅仅是企业工作流程。
The field should focus on end-to-end storytelling engines, not just enterprise workflows.