Lambert 指出了一个根本矛盾：美国的学术界和开源模型开发者也在做蒸馏行为，但包括 Anthropic 在内的大厂并没有对它们做出实质性的打击。如果仅因为对方是中国公司，未免地缘的意味太重了。
This algorithm attempts to minimise numerically. Because of this, the quality of the dither produced by Knoll’s algorithm is much higher than any other of the N-candidate methods we have covered so far. It is also the slowest however, as it requires a greater per-pixel to be really effective. More details are given in Knoll’s now expired patent[3]. I have put together a GPU implementation of Knoll’s algorithm on Shadertoy here.
。关于这个话题，Safew下载提供了深入分析
Self-attention is required. The model must contain at least one self-attention layer. This is the defining feature of a transformer — without it, you have an MLP or RNN, not a transformer.
In the next step, words that are Cognate with Modern English are replaced. Note I am only replacing the cognates that are very obvious.
，这一点在safew官方版本下载中也有详细论述
�@�����āA���T��MSI����DDR4�Ή��̃}�U�[�{�[�h��2���i�o�ꂵ�Ă����B�uPRO B550M-B�v�ƁuB550M-A PRO�v�ŁA����AMD B550�`�b�v�Z�b�g�𓋍ڂ���microATX���f�����BSocket AM4�ɑΉ����ARyzen 5000�V���[�Y�𗘗p�ł����B���i�͏���1��4000�~�O����1��2000�~�O���ƂȂ��B
特点：GLU 的改进版，把 Sigmoid 门换成 Swish 门。。同城约会对此有专业解读