没有生态系统的前沿是不稳定的 | sn scratchpad

A frontier without an ecosystem is not stable

摘要 · Abstract

关于技术进步和现实世界影响的笔记。

Notes on advances in technology and real-world impact

提出“Token 资本”概念，即企业拥有的 AI 能力，与人力资本复合增长。
Introduces 'token capital' as firm-owned AI capability that compounds with human capital.
主张人类能动性驱动 Token 资本增长；任务可外包，学习不可外包。
Argues human agency drives token capital growth; offloading tasks never offloads learning.
提出学习循环架构，通过智能体系统使人力与 Token 资本复合增长。
Proposes a learning loop architecture where human and token capital compound via agentic systems.
强调使用私有评估和强化学习环境，让模型针对业务特定结果改进。
Emphasizes private evals and RL environments for models to improve on business-specific outcomes.
警告价值集中于少数前沿模型的风险，倡导构建前沿生态系统。
Warns against value concentration in few frontier models, advocating a frontier ecosystem.
声称早期构建此学习循环的企业将获得难以复制的优势。
Claims firms that build this learning loop early gain hard-to-replicate advantage.

缺乏实证证据；Token 资本和学习循环等概念未经验证。
Lacks empirical evidence; concepts like token capital and learning loop are not validated.
假设企业能构建专有 AI 系统而不依赖外部模型，可能不现实。
Assumes firms can build proprietary AI systems without relying on external models.
未解决将人类判断融入 AI 循环的技术挑战。
Does not address technical challenges of integrating human judgment into AI loops.
忽视了 AI 可能取代而非增强人类角色的可能性。
Overlooks potential for AI to replace human roles, not just augment them.
该愿景可能仅限于有资源构建此类系统的大型企业。
The vision may be limited to large firms with resources to build such systems.