<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>Richard Jeevash: research and notes</title>
  <link href="https://richardjeevash.com/"/>
  <link rel="self" href="https://richardjeevash.com/feed.xml"/>
  <id>https://richardjeevash.com/</id>
  <updated>2026-11-12T00:00:00Z</updated>
  <author><name>Richard Jeevash</name></author>
  <entry>
    <title>Small decisions, mostly separate parts: how small language models indent code, choose ordinal suffixes and write in capitals</title>
    <link href="https://richardjeevash.com/research-small-decisions.html"/>
    <id>https://richardjeevash.com/research-small-decisions.html</id>
    <published>2026-11-12T00:00:00Z</published>
    <updated>2026-11-12T00:00:00Z</updated>
    <summary>Three small decisions, mostly separate mechanisms: a few late heads copy the indent of the block being closed, an early head and a late MLP apply the “teen, so th” rule, and capitals are added late along one direction. Two late heads have a small second role; the groups that do most of each task’s work are separate.</summary>
  </entry>
  <entry>
    <title>Counting by dilution and a width stored at the newline: a partial account of line-breaking in Pythia-410M</title>
    <link href="https://richardjeevash.com/research-line-breaking.html"/>
    <id>https://richardjeevash.com/research-line-breaking.html</id>
    <published>2026-11-10T00:00:00Z</published>
    <updated>2026-11-10T00:00:00Z</updated>
    <summary>Pythia-410M counts how full a line is by how thin one head’s attention to the line start gets, and reads the last line’s length back from the newline. A partial account: under the stricter test, even the fuller circuit’s fidelity to the model’s decision is only 0.22.</summary>
  </entry>
  <entry>
    <title>Why I post findings before peer review</title>
    <link href="https://richardjeevash.com/notes-posting-before-peer-review.html"/>
    <id>https://richardjeevash.com/notes-posting-before-peer-review.html</id>
    <published>2026-11-04T00:00:00Z</published>
    <updated>2026-11-04T00:00:00Z</updated>
    <summary>Interpretability moves fast and much of it isn’t settled. Posting checked findings early, clearly labelled, seems more useful than holding them back for a year.</summary>
  </entry>
  <entry>
    <title>How I run a study</title>
    <link href="https://richardjeevash.com/notes-how-i-run-a-study.html"/>
    <id>https://richardjeevash.com/notes-how-i-run-a-study.html</id>
    <published>2026-11-04T00:00:00Z</published>
    <updated>2026-11-04T00:00:00Z</updated>
    <summary>Freeze the plan, save every output, generate every number from what was saved, and report what didn’t work. The routine behind each paper on this site.</summary>
  </entry>
</feed>
