Boris 分享他自去年 11 月起 100%用 Claude 在手机上写代码、使用 80 亿 Token 的经历，并讨论企业应如何平衡 AI 模型成本与 ROI，建议给全员发放 Token 进行实验。
Boris shares his experience writing 100% of code with Claude on phone, using 8B tokens since March, and discusses how companies should balance AI model cost vs ROI by giving everyone tokens to experiment.
要点 · TL;DR
ROI 而非成本才是衡量 AI 编码工具的正确指标。 ROI, not cost, is the right metric for AI coding tools.
给每个人 token 去实验，创新往往来自意想不到的角色。 Give everyone tokens to experiment; innovation comes from unexpected roles.
一旦编码问题解决，瓶颈就转移到代码审查、安全性和创意生成。 Once coding is solved, bottlenecks shift to code review, security, and idea generation.
核心观点 · Key points
ROI（投资回报率）才是衡量 AI 编码工具的正确框架，而不仅仅是成本。 ROI is the right framing for AI coding tools, not just cost.
给所有人发放 token 进行实验；创新往往来自意想不到的岗位。 Give everyone tokens to experiment; innovation often comes from unexpected roles.
一旦编码问题被解决，下一个瓶颈就是代码审查、安全性和创意生成。 Once coding is solved, the next bottlenecks are code review, security, and idea generation.
循环（Loops）是智能体之后的下一个抽象层，类似于高阶函数。 Loops are the next abstraction layer after agents, like higher-order functions.
使用最强大的模型，专注于提升回报，而不是削减成本。 Use the most capable model and focus on increasing returns, not cutting costs.
反共识 · Contrarian takes
自 Opus 4.5 以来，Boris 100%的代码由 Claude 编写，而且他在手机上编程。 100% of Boris's code has been written by Claude since Opus 4.5, and he codes on his phone.
自动模式去除了人工审批提示，通过消除提示疲劳来提高安全性。 Auto mode removes human approval prompts, improving security by eliminating prompt fatigue.
Fable 的飞跃比 Opus 4.5 更大；Boris 已经找不到难题来测试它了。 Fable's leap is larger than Opus 4.5; Boris ran out of hard problems to give it.
Claude 安全现在能发现人类渗透测试人员遗漏的问题。 Claude Security now catches issues that human pen testers miss.
动态工作流编排数千个子智能体，是一种新的测试时计算形式。 Dynamic workflows orchestrate thousands of sub-agents, a new form of test-time compute.