Jev 与三年前 ALIGN 文本对齐工作的思路对照

YuhengZha @yzha_zha原推发布于 2026.09.19

在 X 上打开原推

💡Jev reminds me of the text alignment work (ALIGN) we built 3 years ago. Many core ideas are similar: * Next-token prediction is general but may not be efficient for many downstream tasks. * A reusable discriminative function supports different sub-tasks that balance generalization and efficiency. * Collaboration between large generator (e.g., LLM) and small judge (e.g., ALIGN) improves the system's overall reliability.

这条是引用着另一条写的：「ChatGPT 联合发明人发布新模型 Jev」