Cognition 联合创始人、AI 编程助手 Devon 的创造者 Scott Wu 分享了他从数学竞赛到构建尖端 AI 的旅程，包括他第一次喝啤酒、心算技巧以及从高中和哈佛辍学的经历。
Scott Wu, co-founder of Cognition and creator of AI coding agent Devon, shares his journey from math competitions to building cutting-edge AI, including his first beer, mental math tricks, and dropping out of high school and Harvard.
要点 · TL;DR
AI 编程智能体将把工程师从写代码转向高层决策。 AI coding agents will shift engineers from writing code to high-level decision-making.
软件工程需求将因杰文斯悖论而增加，而非减少。 Software engineering demand will increase due to Jevons paradox, not decrease.
企业采用 AI 智能体需要信任和适当权限，而非完全自主。 Enterprise adoption of AI agents requires trust and proper permissions, not full autonomy.
核心观点 · Key points
像 Devin 这样的 AI 编程智能体将把工程师从写代码转向高层决策。 AI coding agents like Devin will shift engineers from writing code to high-level decision-making.
AI 的价值在于整个技术栈中存在有意义差异化的地方。 The value in AI accrues where there is meaningful differentiation across the stack.
由于杰文斯悖论，软件工程需求将增加，而非减少。 Software engineering demand will increase due to Jevons paradox, not decrease.
企业采用 AI 智能体需要信任和适当的权限，而非完全自主。 Enterprise adoption of AI agents requires trust and proper permissions, not full autonomy.
软件工程的未来包括同步（IDE）和异步（智能体）两种体验。 The future of software engineering involves both synchronous (IDE) and asynchronous (agent) experiences.
反共识 · Contrarian takes
成为创始人变得更难是因为行业成熟和现有经验手册，而非竞争减少。 Being a founder has gotten harder due to increased maturity and playbooks, not less competition.
AI 不会消灭软件工程师；它将增加其数量并改变其角色。 AI will not eliminate software engineers; it will increase their numbers and change their role.
AGI 不是一个突然的事件；进步将是渐进的，持续改进。 AGI is not a sudden event; progress will be gradual with continuous improvements.
产品创新落后于 AI 能力；冻结模型仍将带来十年的产品进步。 Product innovation lags behind AI capabilities; freezing models would still yield a decade of product progress.
AI 技术栈的所有层（数据中心、实验室、应用）都将表现良好，而非仅一层。 All layers of the AI stack (datacenter, labs, applications) will do well, not just one.
本期章节 · Chapters（共 26）
开场闲聊与啤酒Opening banter and beer
数学竞赛背景Math competition background
AI 时代的年轻创始人Young founders in AI era
游戏与认知类比Gaming and cognition analogy
软件工程：本质与偶然复杂度Software Engineering: Essential vs Accidental Complexity
企业采用与权限管理Enterprise Adoption and Permissions
衡量 AI 编码的生产力影响Measuring Productivity Impact of AI Coding
编码工具与模型性能Coding tools and model performance
基准任务与模型性能Benchmark tasks and model performance
AI 行业分层与价值获取AI industry layers and value accrual
AI 代理的经济基础设施Economic infrastructure for AI agents
定价模式：按席位 vs. 按用量Pricing Models: Seat-Based vs. Usage-Based
代理经济The Agent Economy
代理的现实障碍Real-World Blockers for Agents
消费应用潜力Consumer App Potential
代理与航空公司谈判Agent Negotiating with Airlines
AI 的信任与经济基础设施Trust and Economic Infrastructure for AI
代理权限与雇佣Agent permissions and hiring
何时雇佣最后一名工程师？When will you hire your last engineer?
AI 用户界面与产品演进AI User Interfaces and Product Evolution
收购时间线与决策Acquisition timeline and decision
AI 领域的许可交易与并购Licensing deals and M&A in AI
Cognition 的文化与收购要约Cognition's culture and buyout offer