[架构提案] 适用于递归自我改进（RSI）的硬件门控遏制框架（Génesis 协议 V5.0）
作者: guienjavar-a11y创建于 2026年9月13日更新于 2026年9月13日
- Problem Statement: Limitations of Software Alignment Currently, software-based alignment methods (such as RLHF, Constitutional AI) are mathematically and architecturally insufficient to contain models with the ability for recursive self-improvement (RSI). Semantic fences are easily overwritten autonomously after the emergence of AGI. In addition, geopolitical competition makes voluntary global pauses impractical. The industry needs a deterministic, hardware-based "fail-safe" architecture that ensures absolute containment while not stopping cutting-edge development.
内容来源: deepseek-ai/DeepSpec