<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Blog on Stephen Sinclair</title>
    <link>https://sinclairs.gitlab.io/blog/</link>
    <description>Recent content in Blog on Stephen Sinclair</description>
    <generator>Hugo</generator>
    <language>en-US</language>
    <lastBuildDate>Thu, 30 Apr 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://sinclairs.gitlab.io/blog/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>My calculator is a transformer</title>
      <link>https://sinclairs.gitlab.io/blog/my-calculator-is-a-transformer/</link>
      <pubDate>Thu, 30 Apr 2026 00:00:00 +0000</pubDate>
      <guid>https://sinclairs.gitlab.io/blog/my-calculator-is-a-transformer/</guid>
      <description>&lt;blockquote&gt;&#xA;&lt;p&gt;TLDR: Article describes a method to define &amp;ldquo;registers&amp;rdquo;, &amp;ldquo;heads&amp;rdquo; and non-linear &amp;ldquo;compute&amp;rdquo;&#xA;functions which are compiled and distilled into a standard Transformer model. This is&#xA;used to develop a simple and deterministic RPN interpreter.&lt;/p&gt;&#xA;&lt;p&gt;Code accompanying this project can be&#xA;&lt;a href=&#34;https://github.com/radarsat1/rpn_transformer&#34;&gt;found here&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;AI disclaimer: code ≈ 50%, non-agentic. Article writing = 0%.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;A few weeks ago I read a blog post from Percepta.ai entitled, &lt;a href=&#34;https://www.percepta.ai/blog/can-llms-be-computers&#34;&gt;Can LLMs be&#xA;computers?&lt;/a&gt;&lt;sup id=&#34;fnref:1&#34;&gt;&lt;a href=&#34;#fn:1&#34; class=&#34;footnote-ref&#34; role=&#34;doc-noteref&#34;&gt;1&lt;/a&gt;&lt;/sup&gt;. The article frames&#xA;this as an alternative to tool calling. Instead of writing out a program and using an&#xA;external tool to evaluate it, just build an exact, deterministic interpreter right into&#xA;the Transformer network. This framing led to some criticism regarding the approach,&#xA;particularly &amp;ldquo;what is the point?&amp;rdquo; and &amp;ldquo;where are the benchmarks?&amp;rdquo; These are fair&#xA;questions, but I think it unfortunately detracted from what I thought was a rather&#xA;fascinating topic, which is: how &lt;em&gt;do&lt;/em&gt; you build a deterministic interpreter into a&#xA;Transformer?&lt;/p&gt;</description>
    </item>
    <item>
      <title>Revsiting the Decision Transformer with Tic Tac Toe</title>
      <link>https://sinclairs.gitlab.io/blog/revsiting-the-decision-transformer-with-tic-tac-toe/</link>
      <pubDate>Tue, 25 Nov 2025 00:00:00 +0000</pubDate>
      <guid>https://sinclairs.gitlab.io/blog/revsiting-the-decision-transformer-with-tic-tac-toe/</guid>
      <description>&lt;blockquote&gt;&#xA;&lt;p&gt;Note: code accompanying this project can be &lt;a href=&#34;https://github.com/radarsat1/dt-ttt&#34;&gt;found&#xA;here&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;This weekend, as an afterthought, I guess due to some background threads that had been&#xA;accumulating in my head for a while, I decided to finally take a quick &amp;amp; dirty shot at one&#xA;of those things that&amp;rsquo;s fascinated me, but not quite enough to actually get around to&#xA;trying it until quite some time has passed: the &lt;strong&gt;Decision Transformer&lt;/strong&gt;&lt;sup id=&#34;fnref:1&#34;&gt;&lt;a href=&#34;#fn:1&#34; class=&#34;footnote-ref&#34; role=&#34;doc-noteref&#34;&gt;1&lt;/a&gt;&lt;/sup&gt; (DT). I&#xA;remember reading about this when it came out and just being a bit.. I dunno,&#xA;simultaneously amazed and annoyed. I had recently finished a round of studying&#xA;reinforcement learning again, and at the time it seemed like people were just throwing&#xA;transformers at everything to see what sticks. At the same time, the idea seemed to just&#xA;make sense: what if we just.. condition a sequence prediction task on how well we want it&#xA;to perform? But it &lt;em&gt;also&lt;/em&gt; completely subverted everything I understood about RL and the&#xA;importance of Bellman value estimation. Anyways, I left it, just curious to see what would&#xA;follow it up, and I guess I didn&amp;rsquo;t hear much more, although there was certainly some&#xA;related work being done.&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
