📄 中文摘要
Nicholas Carlini花费20,000美元和2,000个会话,使用16个Claude实例构建了一个C编译器,该编译器能够编译真实程序。Carlini的关键见解并不在于模型本身,而是在于围绕Claude设计的环境。他的主要努力集中在测试、容器、反馈循环以及基础设施上,这些都能判断代理输出的正确性,因为代理无法自行判断。Chris Lattner对该编译器进行了评审,进一步验证了其有效性。
📄 English Summary
The environment is the product
Nicholas Carlini invested $20,000 and 2,000 sessions to build a C compiler using 16 instances of Claude, successfully compiling real programs. His key insight was not about the model itself but rather the environment designed around Claude. Most of his effort was dedicated to aspects such as testing, containers, feedback loops, and the infrastructure that determines whether the agent's output is correct, as the agent cannot assess this on its own. Chris Lattner reviewed the compiler, further validating its effectiveness.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等