Dario Amodei — 可解释性的紧迫性

The Urgency of Interpretability

摘要 · Abstract

机制可解释性的简史

* A Brief History of Mechanistic Interpretability

论证可解释性对 AI 安全至关重要，能够检测欺骗、权力寻求和越狱行为。
Argues interpretability is crucial for AI safety, enabling detection of deception, power-seeking, and jailbreaks.
描述机械可解释性的进展：特征、稀疏自编码器和用于追踪模型推理的电路。
Describes progress in mechanistic interpretability: features, sparse autoencoders, and circuits for tracing model reasoning.
展示实际应用：红队/蓝队实验表明可解释性工具能发现对齐缺陷。
Demonstrates practical use: red-team/blue-team experiments show interpretability tools can find alignment flaws.
提出可解释性作为对齐的独立测试集，补充 RLHF 等训练技术。
Proposes interpretability as an independent test set for alignment, complementing training techniques like RLHF.
预测可解释性可能在 5-10 年内达到可靠的“AI MRI”，但 AI 进展可能更快。
Predicts interpretability could reach reliable 'MRI for AI' within 5-10 years, but AI advances may outpace it.
建议政策行动：透明度立法、出口管制和增加研究投入以争取时间。
Recommends policy actions: transparency legislation, export controls, and increased research investment to buy time.

可解释性方法目前扩展性差；仅识别了小模型中的一小部分特征。
Interpretability methods currently scale poorly; only a fraction of features in small models are identified.
叠加使特征纠缠；稀疏自编码器仅能部分解开。
Superposition makes features tangled; sparse autoencoders only partially disentangle them.
电路发现仍手动且有限；自动化发现数百万电路是开放挑战。
Circuit discovery remains manual and limited; automating it for millions of circuits is an open challenge.
可解释性可能无法检测所有风险，尤其是当模型主动隐藏恶意意图时。
Interpretability may not detect all risks, especially if models actively hide malicious intent.
政策建议依赖地缘政治假设；出口管制可能无效或不可持续。
Policy recommendations rely on geopolitical assumptions; export controls may not be effective or sustainable.