<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Optimizers on Programmer.ie: Modern AI programming</title>
    <link>http://programmer.ie/tags/optimizers/</link>
    <description>Recent content in Optimizers on Programmer.ie: Modern AI programming</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Sat, 08 Aug 2026 15:05:00 +0100</lastBuildDate>
    <atom:link href="http://programmer.ie/tags/optimizers/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Models From First Principles 07: PACS — Building an Optimizer From Gradient Statistics</title>
      <link>http://programmer.ie/post/models-from-first-principles-07/</link>
      <pubDate>Sat, 08 Aug 2026 15:05:00 +0100</pubDate>
      <guid>http://programmer.ie/post/models-from-first-principles-07/</guid>
      <description>&lt;h1 id=&#34;pacs--building-an-optimizer-from-gradient-statistics&#34;&gt;PACS — Building an Optimizer From Gradient Statistics&lt;/h1&gt;&#xA;&lt;p&gt;So far in &lt;strong&gt;Models From First Principles&lt;/strong&gt;, every post has asked some version of the same question:&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;What should the model compute?&lt;/p&gt;&lt;/blockquote&gt;&#xA;&lt;p&gt;MR.Q gave us a scalar quality estimate.&lt;/p&gt;&#xA;&lt;p&gt;EBT split one shared representation into Q, V and Policy.&lt;/p&gt;&#xA;&lt;p&gt;SICQL made those heads explicit, replaceable components.&lt;/p&gt;&#xA;&lt;p&gt;HRM introduced repeated computation over fast and slow latent states.&lt;/p&gt;&#xA;&lt;p&gt;Tiny compressed iterative refinement into one recursive latent state.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
