Unverified50% confidenceBenchmarkExact time
Anthropic招募了1053名程序员做对照实验，在正常任务中混入危险命令，人类放行了800条危险命令，机器只漏掉了6条
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified工程师Boris让AI智能体在数周内自动执行模糊测试与死代码清理，累计提交了388个PR，其中180个已成功合入主分支64% similarUnverifiedAFK智能体在约一个半小时内无人值守完成了5轮迭代，生成了6次附带详细提交信息的代码提交60% similarUnverifiedMigrating 750,000 lines of code by a human team would require several hundred engineers working in parallel60% similarUnverifiedAnthropic推出动态工作流功能，Jard Sumner使用该功能将Bun从Zig移植到Rust，生成约75万行Rust代码，测试套件通过率达99.8%，从首次提交到合并约11天60% similarUnverified纳瓦尔·拉维坎特声称自己一个人用氛围编程重做了当年需要8个工程师干一年才能完成的App60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/897099API
curl https://kongchang.com/api/v1/knowledge/claims/897099MCP
get_claim(id=897099)