<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Tool-Use on Notes from the Rabbit Hole</title><link>https://magnus919.com/tags/tool-use/</link><description>Recent content in Tool-Use on Notes from the Rabbit Hole</description><generator>Hugo</generator><language>en</language><copyright>© [Magnus Hedemark](https://github.com/magnus919)</copyright><lastBuildDate>Thu, 09 Jul 2026 21:30:15 -0400</lastBuildDate><atom:link href="https://magnus919.com/tags/tool-use/index.xml" rel="self" type="application/rss+xml"/><item><title>Andrej Karpathy's Cognitive Core: The Model Is Not the Knowledge</title><link>https://magnus919.com/2026/07/andrej-karpathys-cognitive-core-the-model-is-not-the-knowledge/</link><pubDate>Thu, 09 Jul 2026 21:30:15 -0400</pubDate><guid>https://magnus919.com/2026/07/andrej-karpathys-cognitive-core-the-model-is-not-the-knowledge/</guid><description>&lt;p>What if the most important AI development isn&amp;rsquo;t a model that knows more, but one that knows less?&lt;/p>
&lt;p>That is the question at the heart of Andrej Karpathy&amp;rsquo;s &amp;ldquo;cognitive core&amp;rdquo; thesis. In a &lt;a href="https://x.com/karpathy/status/1938626382248149433">tweet&lt;/a> and a &lt;a href="https://www.dwarkesh.com/p/andrej-karpathy">podcast interview&lt;/a>, he laid out a vision that sounds almost heretical in an industry built on scaling: build a small model (a few billion parameters) that intentionally sacrifices encyclopedic knowledge for raw reasoning capability. A model that lives always-on on your computer, the way an operating system kernel does. It doesn&amp;rsquo;t know when William the Conqueror&amp;rsquo;s reign ended, but it vaguely recognizes the name and can look it up. It can&amp;rsquo;t recite the SHA-256 of an empty string, but it can calculate it if you ask.&lt;/p></description></item></channel></rss>