<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>negmbots lab notes</title><link>https://blog.negmbots.com/</link><description>Recent content on negmbots lab notes</description><generator>Hugo</generator><language>en</language><atom:link href="https://blog.negmbots.com/index.xml" rel="self" type="application/rss+xml"/><item><title>About</title><link>https://blog.negmbots.com/about/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://blog.negmbots.com/about/</guid><description>&lt;h2 id="what-this-blog-does"&gt;What this blog does&lt;/h2&gt;&#10;&lt;p&gt;Once a month I take a recent paper on serving, observability or efficiency of AI models and try it for real:&lt;/p&gt;&#10;&lt;ol&gt;&#10;&lt;li&gt;&lt;strong&gt;Reproduce&lt;/strong&gt; the paper&amp;rsquo;s main claim with the authors&amp;rsquo; released code.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Extend&lt;/strong&gt; it with one question the paper leaves open.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Report&lt;/strong&gt; what held up, what didn&amp;rsquo;t, and what it cost.&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;p&gt;Results are reported as measured, including failures. When I speculate, I say so.&lt;/p&gt;&#10;&lt;h2 id="the-lab"&gt;The lab&lt;/h2&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;Raspberry Pis of several generations, for edge and thermal experiments&lt;/li&gt;&#10;&lt;li&gt;Intel NUCs (i7 and i5) running Proxmox, for CPU inference and the monitoring stack&lt;/li&gt;&#10;&lt;li&gt;GMKtec EVO-X2 with 128 GB of unified memory, for large models on AMD hardware&lt;/li&gt;&#10;&lt;li&gt;Google Cloud GPUs and TPUs, used briefly and with hard cost limits&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;p&gt;The full setup is described in &lt;a href="https://blog.negmbots.com/posts/2026-11-my-lab/"&gt;the first post&lt;/a&gt;.&lt;/p&gt;</description></item></channel></rss>