<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Model Alignment on Matt Suiche</title><link>https://www.msuiche.com/tags/model-alignment/</link><description>Recent content in Model Alignment on Matt Suiche</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Sun, 16 Aug 2026 00:00:00 +0200</lastBuildDate><atom:link href="https://www.msuiche.com/tags/model-alignment/index.xml" rel="self" type="application/rss+xml"/><item><title>Autoresearch: Shipping a Behaviour Change in 20 Kilobytes</title><link>https://www.msuiche.com/posts/autoresearch-shipping-a-behaviour-change-in-20-kilobytes/</link><pubDate>Sun, 16 Aug 2026 00:00:00 +0200</pubDate><guid>https://www.msuiche.com/posts/autoresearch-shipping-a-behaviour-change-in-20-kilobytes/</guid><description>&lt;p&gt;Changing what a model refuses usually means redistributing the model. You edit a few
hundred matrices, re-upload 157 gigabytes, and every user pulls a fresh copy of a
checkpoint that differs from the old one by a rounding error smeared thinly across
its weights.&lt;/p&gt;
&lt;p&gt;There&amp;rsquo;s a second option that has been available the whole time. Ship the &lt;em&gt;difference&lt;/em&gt;
as a single vector (about 20 kilobytes of floats) applied at inference. The base
checkpoint stays byte-identical and already cached.&lt;/p&gt;</description></item></channel></rss>