https://www.msuiche.com/tags/weight-svd/Recent content in Weight SVD on Matt SuicheHugoen-usThu, 29 Oct 2026 00:00:00 +0100https://www.msuiche.com/posts/glp-spec-3-steering-deepseek-v4.1-at-the-fold/Thu, 29 Oct 2026 00:00:00 +0100https://www.msuiche.com/posts/glp-spec-3-steering-deepseek-v4.1-at-the-fold/<p>GLP vectors hook one site: the residual stream after a layer&rsquo;s writes have folded in. Every vector shipped before today - DeepSeek V4, Qwen, GLM, Llama-family - names that same tensor, because on those architectures there is exactly one stream and it sits plainly between layers. DeepSeek V4.1 removes it. The model carries each token as four parallel hyper-connection copies, and a layer&rsquo;s FFN write waits a full layer before entering the stream, through a gated mixer DeepSeek calls <code>mhc_post</code>. The old hook name refers to no tensor a probe can tap on this architecture.</p>