<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>safety on Sparse Notes</title>
    <link>https://sparsenotes.com/tags/safety/</link>
    <description>Recent content in safety on Sparse Notes</description>
    <image>
      <url>https://sparsenotes.com/images/og-default.png</url>
      <link>https://sparsenotes.com/images/og-default.png</link>
    </image>
    <generator>Hugo -- gohugo.io</generator>
    <lastBuildDate>Sat, 15 Aug 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://sparsenotes.com/tags/safety/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Synthetic Persona Pretraining: Alignment from Token Zero</title>
      <link>https://sparsenotes.com/posts/2026/08/papers/synthetic-persona-pretraining-alignment-token-zero/</link>
      <pubDate>Sat, 15 Aug 2026 00:00:00 +0000</pubDate>
      
      <guid>https://sparsenotes.com/posts/2026/08/papers/synthetic-persona-pretraining-alignment-token-zero/</guid>
      <description>Annotating 10% of the pretraining corpus with first-person value reflections installs the assistant persona from step zero — and the alignment advantage over post-training-only grows with compute.</description>
    </item>
    
  </channel>
</rss>
