卡帕西描述了一种由“编程”向“意志体现”转变的范式跃迁，其个人状态表现为每日约16小时高强度人机交互。关键量化变化体现在工作分配从历史上的80%人工编码、20%辅助，转向接近0%人工编码、100%由智能体执行的极端结构。这一0/100转变将生产力瓶颈从模型能力转移至人类“技能问题”，即指令质量、记忆结构与Token吞吐管理能力。结果是开发者角色从执行者转为指挥者，能力差距呈指数放大，低效个体被系统性淘汰。
在操作层面，工作模式演化为“多线程”智能体协同，单一界面可并行数十个会话，实现从函数级到代码库级的尺度跃迁。任务被拆分为并行指令流，如重构、逆向工程与规划同步执行，整体效率呈倍数级提升。现实应用如Dobby系统展示了端到端自动化能力：自主扫描网络、识别设备、完成API逆向并控制多系统。该路径预示GUI需求下降，交互由点击转向自然语言指令，软件用户从人类转为智能体中介。
在研究与认知层面，AutoResearch通过递归优化实现AI训练AI，消除人类瓶颈。实验显示，即使具备20年经验的专家也难以发现如权重衰减与Adam参数耦合等细节，而智能体可通过持续试验优化。智能表现呈“锯齿状”：在可验证任务中因强化学习具备明确奖励信号而超人类，在不可验证领域如幽默仍停留低水平。教育体系亦随之重构，文档从面向人类转向面向智能体，知识传递经由中介完成。最终，资源衡量从货币转向算力，核心指标为FLOPs规模与可调度智能体数量。
Karpathy describes a paradigm shift from “programming” to “manifestation,” marked by roughly 16 hours of daily high-intensity human–AI interaction. The key quantitative change is the transition from a historical 80% human coding and 20% assistance split to an about 0% human coding and 100% agent execution structure. This 0/100 inversion relocates the productivity bottleneck from model capability to human “skill issues,” specifically instruction quality, memory design, and token throughput management. Developers shift from executors to commanders, with capability gaps amplifying exponentially and low-efficiency individuals systematically eliminated.
Operationally, workflows evolve into multi-threaded agent collaboration, with dozens of parallel sessions enabling a scale jump from function-level to codebase-level control. Tasks decompose into concurrent streams such as refactoring, reverse engineering, and planning, producing multiplicative efficiency gains. Systems like Dobby demonstrate end-to-end autonomy: scanning networks, identifying devices, reverse-engineering APIs, and controlling environments. This trajectory implies declining GUI relevance, with interaction shifting from clicks to natural language, and software users transitioning from humans to agent intermediaries.
In research, AutoResearch applies recursive self-improvement where AI trains AI, removing human bottlenecks. Evidence shows even experts with 20 years of experience miss micro-optimizations such as weight decay in value embeddings and Adam parameter coupling, while agents discover them through continuous experimentation. Intelligence exhibits “jaggedness”: superhuman performance in verifiable domains with clear reinforcement signals, but weak performance in unverifiable domains like humor. Education restructures accordingly, with documentation targeting agents rather than humans. Ultimately, value metrics shift from currency to compute, defined by FLOPs capacity and controllable agent scale.