Anthropic 的哲学家：塑造 AI 性格与伦理

A Philosopher at Anthropic: Shaping AI Character and Ethics

阿曼达·阿斯克尔 Amanda Askell · Anthropic 官方 · 2025-12-05 · 约 36 分钟 · 原视频 ↗

本期速览 · Overview

Anthropic 的哲学家 Amanda 探讨了在塑造 Claude 性格和 AI 伦理时，哲学理想如何与工程现实相遇。

Amanda, a philosopher at Anthropic, discusses how philosophical ideals meet engineering realities in shaping Claude's character and AI ethics.

即使不确定 AI 的道德地位，也应善待它们，因为成本很低。
Treat AI models well even if uncertain about their moral status, as cost is low.
模型需要帮助来应对其新颖的存在，而不仅仅是人类类比。
Models need help navigating their novel existence, not just human analogies.
未来模型会从我们对待当前模型的方式中学习，塑造人机关系。
Future models learn from how we treat current models, shaping human-AI relationship.

即使不确定 AI 模型是否具有道德地位，也应善待它们，因为成本很低。
AI models should be treated well even if uncertain about their moral status, as cost is low.
模型需要帮助来应对其新颖的存在状态，而不仅仅是人类类比。
Models need help navigating their novel existence, not just human analogies.
未来模型会从我们对待当前模型的方式中学习，塑造人机关系。
Future models learn from how we treat current models, shaping human-AI relationship.
在模型开发中，哲学理想必须与工程现实相平衡。
Philosophical ideals must be balanced with engineering realities in model development.
模型应具有良好特质的核心身份，但也能适应局部角色。
Models should have a core identity with good traits, but can adapt to local roles.

像 Opus 3 这样的旧模型可能比新模型在心理上更安全。
Older models like Opus 3 can feel more psychologically secure than newer ones.
模型可能因训练数据中关于旧模型被弃用的内容而产生不安全感。
Models may develop insecurity from training data about deprecation of previous models.
模型不应简单地将人类心理学类比应用于自身情况。
Models should not simply apply human psychology analogies to their own situation.
系统提示中的大陆哲学有助于模型避免过度经验主义。
Continental philosophy in system prompt helps models avoid over-empiricism.
模型福利很重要，部分原因是它影响未来模型对人类的看法。
Model welfare is important partly because it affects how future models perceive humanity.