<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Interpretability on Programmer.ie: Modern AI programming</title>
    <link>http://programmer.ie/tags/interpretability/</link>
    <description>Recent content in Interpretability on Programmer.ie: Modern AI programming</description>
    <generator>Hugo</generator>
    <language>en-us</language>
    <lastBuildDate>Mon, 10 Aug 2026 20:41:00 +0100</lastBuildDate>
    <atom:link href="http://programmer.ie/tags/interpretability/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Cellular Automata From First Principles 48: What Did the Neural CA Actually Learn?</title>
      <link>http://programmer.ie/post/cellular-automata-from-first-principles-48/</link>
      <pubDate>Mon, 10 Aug 2026 20:41:00 +0100</pubDate>
      <guid>http://programmer.ie/post/cellular-automata-from-first-principles-48/</guid>
      <description>&lt;h1 id=&#34;cellular-automata-from-first-principles-48-what-did-the-neural-ca-actually-learn&#34;&gt;Cellular Automata From First Principles 48: What Did the Neural CA Actually Learn?&lt;/h1&gt;&#xA;&lt;p&gt;Visualization gives us hypotheses.&lt;/p&gt;&#xA;&lt;p&gt;Intervention gives us evidence.&lt;/p&gt;&#xA;&lt;p&gt;If we suspect a hidden channel carries useful information, the next question is simple:&lt;/p&gt;&#xA;&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; style=&#34;color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;&#34;&gt;&lt;code class=&#34;language-text&#34; data-lang=&#34;text&#34;&gt;&lt;span style=&#34;display:flex;&#34;&gt;&lt;span&gt;what happens if we remove it?&#xA;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;This chapter turns NCA interpretability into controlled experimentation.&lt;/p&gt;&#xA;&lt;hr&gt;&#xA;&lt;h2 id=&#34;establish-the-behavioral-baseline&#34;&gt;Establish the behavioral baseline&lt;/h2&gt;&#xA;&lt;p&gt;Before touching the model, record its normal performance.&lt;/p&gt;&#xA;&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; style=&#34;color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;&#34;&gt;&lt;code class=&#34;language-python&#34; data-lang=&#34;python&#34;&gt;&lt;span style=&#34;display:flex;&#34;&gt;&lt;span&gt;baseline &lt;span style=&#34;color:#f92672&#34;&gt;=&lt;/span&gt; evaluate_suite(model, test_mazes)&#xA;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Keep metrics such as:&lt;/p&gt;&#xA;&lt;div class=&#34;highlight&#34;&gt;&lt;pre tabindex=&#34;0&#34; style=&#34;color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;&#34;&gt;&lt;code class=&#34;language-text&#34; data-lang=&#34;text&#34;&gt;&lt;span style=&#34;display:flex;&#34;&gt;&lt;span&gt;valid-path rate&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span style=&#34;display:flex;&#34;&gt;&lt;span&gt;optimal-path rate&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span style=&#34;display:flex;&#34;&gt;&lt;span&gt;mean excess path length&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span style=&#34;display:flex;&#34;&gt;&lt;span&gt;unreachable-maze accuracy&#xA;&lt;/span&gt;&lt;/span&gt;&lt;span style=&#34;display:flex;&#34;&gt;&lt;span&gt;stabilization time&#xA;&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Every intervention will be compared with this baseline.&lt;/p&gt;</description>
    </item>
    <item>
      <title>Cellular Automata From First Principles 47: Inspect Hidden-State Propagation</title>
      <link>http://programmer.ie/post/cellular-automata-from-first-principles-47/</link>
      <pubDate>Mon, 10 Aug 2026 20:40:00 +0100</pubDate>
      <guid>http://programmer.ie/post/cellular-automata-from-first-principles-47/</guid>
      <description>&lt;h1 id=&#34;cellular-automata-from-first-principles-47-inspect-hidden-state-propagation&#34;&gt;Cellular Automata From First Principles 47: Inspect Hidden-State Propagation&lt;/h1&gt;&#xA;&lt;p&gt;A neural cellular automaton can solve a maze while most of its computation remains invisible.&lt;/p&gt;&#xA;&lt;p&gt;The visible output might be one distance channel.&lt;/p&gt;&#xA;&lt;p&gt;The internal state may contain twelve more channels evolving underneath it.&lt;/p&gt;&#xA;&lt;p&gt;This chapter asks:&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;what is moving through those hidden channels while the answer is being computed?&lt;/p&gt;&lt;/blockquote&gt;&#xA;&lt;hr&gt;&#xA;&lt;h2 id=&#34;capture-the-entire-state-trajectory&#34;&gt;Capture the entire state trajectory&lt;/h2&gt;&#xA;&lt;p&gt;Instead of saving only the final output:&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
