耗费2万美元、两周写10万行Rust代码！16个Claude智能体写的C编译器，能编译Linux内核却卡在“Hello World”？

Core Insights - The article discusses a groundbreaking experiment conducted by Anthropic researcher Nicholas Carlini, where a team of 16 Claude AI agents autonomously built a Rust-based C compiler capable of compiling the Linux 6.9 kernel without human intervention [1][4][5]. Group 1: Experiment Overview - The experiment lasted approximately two weeks, involving nearly 2000 Claude Code sessions, consuming around 20 billion input tokens and 1.4 million output tokens, with an API cost close to $20,000, resulting in a C compiler with about 100,000 lines of code [4]. - The compiler demonstrated capabilities beyond previous expectations of large language model (LLM) programming abilities, achieving a 99% pass rate on major compiler test suites and successfully compiling and running the game Doom [7][5]. Group 2: Methodology and Innovation - The key innovation of this experiment lies not in the model itself but in the collaborative approach, where the AI agents were set strict goals to work independently without relying on human input [6][9]. - A simple loop framework was established to allow the agents to take on new tasks immediately after completing previous ones, running within Docker containers to prevent local machine impact [6][8]. Group 3: Challenges and Limitations - Despite the impressive outcomes, the compiler faced criticism for not being able to compile a basic "hello world" program without manual intervention, raising questions about its maturity [10][13]. - Carlini explicitly outlined several limitations of the compiler, including its inability to independently compile the Linux kernel, reliance on GCC components for assembly and linking, and lower performance compared to established compilers [14][15][20]. - The project highlighted that the real challenge was not just writing code but creating an environment that allowed the AI to operate autonomously, emphasizing the need for rigorous testing and feedback mechanisms [21].