<?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>Comprehension on Blackwell Systems</title><link>https://blog.blackwell-systems.com/tags/comprehension/</link><description>Recent content in Comprehension on Blackwell Systems</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Wed, 30 Sep 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://blog.blackwell-systems.com/tags/comprehension/index.xml" rel="self" type="application/rss+xml"/><item><title>A Format Can Save Tokens and Still Be Unreadable. We Measured It.</title><link>https://blog.blackwell-systems.com/posts/bpp-comprehension-tokens-you-cant-read/</link><pubDate>Wed, 30 Sep 2026 00:00:00 +0000</pubDate><guid>https://blog.blackwell-systems.com/posts/bpp-comprehension-tokens-you-cant-read/</guid><description>A whitespace-delimited format called bpp beat GCF on token count in an integration&amp;#39;s benchmark. So we ran an adversarial comprehension study: 8 models, 5 families, ~770,000 records read. Fewer tokens, far more wrong answers. GCF 31% error, JSON 40%, bpp 54%. Full data and reproduction.</description></item></channel></rss>