Arbitrary rejection or suspension without clear justification
This started with Addition Under Pressure, where I gave Claude Code and Codex the same prompt: train the smallest possible transformer that can do 10-digit addition with at least 99% accuracy. Claude Code came back with 6,080 parameters and Codex came back with 1,644. The community has since pushed this dramatically lower.
，这一点在51吃瓜中也有详细论述
�@�x�[�X�X�e�[�V�����ɂ�4���̃E�I�[�^�[�W�F�b�g�Ƒ����a�̋z���|�[�g�𓋍ڂ��A���ꂽ���┯�̖сA�傫�ȃS�~�Ȃǂ̏������\�B�ő�200���ԃ����e�i���X�t���[�ŗ��p���邱�Ƃ��\���B
Continue reading...
把 Claude 的输出直接喂给另一个架构的模型，不一定有效，有时甚至会产生干扰。两个模型内部表征空间的差异，会让「老师」的回答在「学生」那里引发意想不到的偏差。