arXiv：从“种子AI”到技术奇点

TL;DR - This article reviews a 2015 arXiv paper on whether “seed AI” could recursively improve its software toward superintelligence. It argues that such progress faces fundamental computational, physical, logical, verification, and safety constraints.

• True recursive self-improvement requires each generation to become better at improving itself, not merely adapt parameters or gain hardware.
• Proposed paths include universal program search, systems that scientifically redesign themselves, and indirect human–AI improvement loops.
• Self-reference, undecidable problems, diminishing returns, accumulated errors, and incomplete self-modeling may prevent unlimited improvement.
• Preserving goals and safety guarantees across major code rewrites remains a central unresolved challenge.