<?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>Scaling on Zero to One AI</title><link>https://blog.crazyai.uk/tags/scaling/</link><description>Recent content in Scaling on Zero to One AI</description><generator>Hugo</generator><language>en</language><copyright>Volume 01 · Zero to One AI · Typeset in public</copyright><lastBuildDate>Sun, 30 Aug 2026 19:50:00 +0800</lastBuildDate><atom:link href="https://blog.crazyai.uk/tags/scaling/index.xml" rel="self" type="application/rss+xml"/><item><title>Scale Bought the Base</title><link>https://blog.crazyai.uk/posts/adapter-field-notes-03-scale-bought-the-base/</link><pubDate>Sun, 30 Aug 2026 19:50:00 +0800</pubDate><guid>https://blog.crazyai.uk/posts/adapter-field-notes-03-scale-bought-the-base/</guid><description>Qwen3-14B was a stronger base than 8B. After the same 50-step QLoRA, the three adapters sat on 108, 106, and 107 GSM8K problems.</description></item></channel></rss>