用 Jev 构建代码库分类器
开发者分享用 Jev 构建代码库分类器,认为这可能解决智能体生成过度工程化代码的问题,并询问下一步测试方向。
作者尝试用 Jev 玩马里奥,但因日本延迟无法成功;于是用 Jev 生成教师数据训练 LightGBM 并替换,实现从 Jev 的 300 倍以上加速并通关。
I tried playing Mario with Jev, but due to latency issues from Japan, no matter how many times I attempted it, I couldn't succeed. However, by creating teacher data with Jev, training LightGBM on it, and replacing it, I was able to speed it up by over 300 times from Jev and clear