<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>ARGMAX (argmaxai.dev) — The 0.1% SOTA Intelligence Engine</title>
    <link>https://argmaxai.dev</link>
    <description>Deterministic mathematical breakdowns, Triton kernel replications, and verified benchmark deltas for principal AI architects, ML teams, and quantitative funds.</description>
    <language>en-us</language>
    <lastBuildDate>Mon, 19 Oct 2026 13:07:37 GMT</lastBuildDate>
    <atom:link href="https://argmaxai.dev/rss.xml" rel="self" type="application/rss+xml"/>
    
    <item>
      <title>Dual-Chunk Recurrent Linear Attention: Sub-Quadratic Scaling at 128k Context</title>
      <link>https://argmaxai.dev/dispatches/dual-chunk-linear-attention</link>
      <guid>https://argmaxai.dev/dispatches/dual-chunk-linear-attention</guid>
      <pubDate>Sun, 18 Oct 2026 08:00:00 GMT</pubDate>
      <description><![CDATA[Replaces quadratic multi-head attention with a dual-chunk recurrent linear attention kernel. 3.2x faster decode speed, reduces KV cache from 18.4GB to 1.9GB on 70B parameter models at 128k context with zero accuracy degradation on SWE-bench Verified.]]></description>
    </item>

    <item>
      <title>DeepSeek-V3 vs Claude 3.7: Multi-Head Latent Attention &amp; FP8 Triton Kernel Deconstruction</title>
      <link>https://argmaxai.dev/dispatches/deepseek-v3-latent-attention-fp8</link>
      <guid>https://argmaxai.dev/dispatches/deepseek-v3-latent-attention-fp8</guid>
      <pubDate>Wed, 14 Oct 2026 08:00:00 GMT</pubDate>
      <description><![CDATA[An exhaustive reverse-engineering of the memory bandwidth bottleneck, Triton kernel update formulations, and independent verification of Table 2 claims.]]></description>
    </item>
  </channel>
</rss>