普林斯顿 CEO-Bench 测试：AI 运营虚拟公司多数破产
普林斯顿用 CEO-Bench 模拟 500 天公司经营，多数 AI 模型破产，说明通用 Agent 需针对垂直场景深度适配。
普林斯顿大学推出 CEO-Bench，让 AI 在 500 天内运营一家虚拟 SaaS 初创。多个前沿模型因定价、营销等决策失误而破产，只有少数如 GPT-5.5 能灵活调整策略。研究发现，通用编程 Agent 直接套用 CEO 角色效果不佳，垂直适配才是关键。
正文摘录
- title: AI as CEO: Quickly Driving 10 Companies to Bankruptcy… - sourcecompany: QbitAI - bodymarkdown: Princeton University recently launched CEO-Bench, letting AI run a virtual SaaS startup for 500 days. The core is a Python API with 34 tools and 19 database tables. Once the Agent connects, it can write code, query databases with SQL, and dynamically adjust workflows based on the query results. Pricing strategy, ad spend channels, R&D budget allocation, infrastructure scaling, customer support team configuration—it all has to figure it out on its own. Customer satisfaction, willingness to pay, minimum quality expectations—these metrics can only be inferred indirectly from churn rates, ticket volumes, and social networks. Meanwhile, the external environment is always changing dynamically:…