arXiv:2511.10668 · 中英对照阅读
人工智能奇点的数学框架:递归改进的条件、界限与控制
A Mathematical Framework for AI Singularity: Conditions, Bounds, and Control of Recursive Improvement
中文速览
AI 能否因“越聪明越会让自己变聪明”而在有限时间内失控增长,关键要看能力提升速度是否会随能力本身加速到超过资源限制。论文把能力、算力、数据、能源和资本放进一个可测量的数学模型,用增长不等式判断何时会爆发,并把功耗、带宽、存储等现实上限,以及限电、限速和评估关卡纳入分析。结果表明,增长速度若呈足够强的超线性加速,有限时间爆发才可能发生;反之,物理资源上限和适当控制可以给出可检验的安全界限。它的意义是把“奇点会不会来”的猜测变成可用基准测试和设施数据检验、也能通过实际管理措施加以约束的问题。
摘要
AI systems improve by using more compute, data, and better algorithms. A common fear is that this improvement might feed on itself and run away—what people call a “singularity” [18]. This paper asks a precise version of that question: under what measurable conditions could capability growth run away in finite time, and under what conditions can we rule that out? We aim for answers that are testable with real power, bandwidth, and benchmark data, and for controls that can be implemented in practice (e.g., power caps, throttling, evaluation gates). The goal is to replace speculation with clear conditions that either certify safety or flag genuine risk—without relying on simulations or unverifiable assumptions.
术语表
- AI singularity
- 人工智能奇点
- finite-time blow-up
- 有限时间爆发
- capability process
- 能力过程
- recursive self-improvement
- 递归自我改进
- improvement operator
- 改进算子
- compute
- 计算资源
- scaling law
- 缩放定律
- algorithmic efficiency
- 算法效率
- Osgood–Bihari inequalities
- Osgood–Bihari 不等式
- comparison principle
- 比较原理
- Lyapunov barrier certificate
- Lyapunov 障碍证书
- blow-up criterion
- 爆发判据
- stabilization
- 稳定化
- identifiability inequality
- 可辨识性不等式
- energy-to-erasure limit
- 能量擦除极限
- memory–communication constraint
- 存储–通信约束
- critical manifold
- 临界流形
- control input
- 控制输入
- forward invariant set
- 正向不变集
- control barrier certificate
- 控制障碍证书
- viability theory
- 可生存性理论
- monotone envelope argument
- 单调包络论证
- empirical scaling law
- 经验缩放定律
- learning-theoretic lower bound
- 学习理论下界
- capital accumulation model
- 资本积累模型