https://www.msuiche.com/tags/speculative-decoding/Recent content in Speculative Decoding on Matt SuicheHugoen-usThu, 03 Sep 2026 00:00:00 +0200https://www.msuiche.com/posts/autoresearch-sticky-refusals-free-speculative-decoding-and-the-invisible-quantisation-cliff/Thu, 03 Sep 2026 00:00:00 +0200https://www.msuiche.com/posts/autoresearch-sticky-refusals-free-speculative-decoding-and-the-invisible-quantisation-cliff/<p>This is the follow-up to the <a href="https://www.msuiche.com/posts/autoresearch-abliteration-without-redistributing-the-model/">projection-steering post</a>: two more weeks, five model families, and a pile of measurements that killed several of my own assumptions. The short version is that the GLP approach, ship the <em>difference</em>, not the model, now covers seven checkpoints from five vendors (DeepSeek, Qwen, Z.ai, Thinking Machines, Tencent), and the interesting findings are no longer &ldquo;it works&rdquo; but <em>where it behaves differently</em>, <em>what it composes with</em>, and, the new thread, <em>what it breaks that isn&rsquo;t refusal</em>.</p>