尽管从理论上讲，随着AI模型需要完成更复杂的任务，强化学习的训练会需要更多FLOPs（浮点运算），从而拖慢进展速度，但实际AI进步并未显著放缓。本文提出了三个原因：新模型正以数量级更高的效率利用现有FLOPs（例如修复像FP16精度错误这样的低级bug）；人类对AI智能水平的判断存在偏差，模型越接近人类水平就越难衡量其真实进步；除了智能之外，工作记忆、持久性（persistence）和个性等多种特质共同决定了模型的实际能力。作者认为，AI发展主要受制于偶发性的突破（如巧妙想法或修复重大bug），而非单纯的理论瓶颈。
Longer training horizons have not slowed AI progress because modern architectures (like transformers) and techniques such as reward shaping distribute gradients effectively, making credit assignment over long episodes tractable without exponential difficulty increases.
Andrew Ng ranks coding agent acceleration from most to least effective: frontend development (dramatically sped up), backend (slower due to bugs and security), infrastructure (limited due to complex tradeoffs), and research (marginal help beyond coding). He uses this categorization to adjust team expectations and organization.