<?xml version="1.0" encoding="utf-8"?>
<?xml-stylesheet href="https://blog.sagamiyun.me/feed_style.xsl" type="text/xsl"?>
<feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en">
    <tabi:metadata xmlns:tabi="https://github.com/welpo/tabi">
        <tabi:base_url>https:&#x2F;&#x2F;blog.sagamiyun.me</tabi:base_url>
        <tabi:separator>
            •
        </tabi:separator>
        <tabi:about_feeds>This is a web feed, also known as an Atom feed. Subscribe by copying the URL from the address bar into your newsreader. Visit About Feeds to learn more and get started. It&#x27;s free.</tabi:about_feeds>
        <tabi:visit_the_site>Visit website</tabi:visit_the_site>
        <tabi:recent_posts>Recent posts</tabi:recent_posts>
        <tabi:last_updated_on>Updated on $DATE</tabi:last_updated_on>
        <tabi:default_theme>dark</tabi:default_theme>
        <tabi:post_listing_date>date</tabi:post_listing_date>
        <tabi:current_section>rl</tabi:current_section>
    </tabi:metadata><link rel="extra-stylesheet" href="https://blog.sagamiyun.me/skins/teal.css?h=bd19e558a52d678a50de" /><title>Kit Kyo - rl</title>
        <subtitle>Full-stack engineer · DevOps · agent development — currently going deep on ML &amp; infrastructure.</subtitle>
    <link href="https://blog.sagamiyun.me/tags/rl/atom.xml" rel="self" type="application/atom+xml"/>
    <link href="https://blog.sagamiyun.me/tags/rl/" rel="alternate" type="text/html"/>
    <generator uri="https://www.getzola.org/">Zola</generator><updated>2026-06-20T00:00:00+00:00</updated><id>https://blog.sagamiyun.me/tags/rl/atom.xml</id><entry xml:lang="en">
        <title>Kill the tail, not the cores: profiling the verifier under RL-with-verifiers</title>
        <published>2026-06-20T00:00:00+00:00</published>
        <updated>2026-06-20T00:00:00+00:00</updated>
        <author>
            <name>Kit Kyo · A2O Labs</name>
        </author>
        <link rel="alternate" href="https://blog.sagamiyun.me/blog/kill-the-tail-not-cores/" type="text/html"/>
        <id>https://blog.sagamiyun.me/blog/kill-the-tail-not-cores/</id>
        <summary type="html">As RL-with-verifiers scales, the Lean verifier is both the reward source and the throughput bottleneck — a CPU&#x2F;scheduling problem almost nobody benchmarks on commodity, on-prem hardware. We profiled it: on an 8-core box, ~4.5% of proofs run away and eat most of the single-thread wall, and a tuned per-proof deadline beats adding cores. The honest-measurer story (including the claim we had to retract) is the point.</summary>
        </entry>
</feed>
