<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Context Engineering | Digitaliziran si</title><link>https://digitaliziran.si/</link><description>Context Engineering | Digitaliziran si</description><generator>Hugo</generator><language>en</language><lastBuildDate>Mon, 26 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://digitaliziran.si/tags/context-engineering/index.xml" rel="self" type="application/rss+xml"/><item><title>The Half-Life of an Instruction</title><link>https://digitaliziran.si/2026/10/26/the-half-life-of-an-instruction.html</link><pubDate>Mon, 26 Oct 2026 00:00:00 +0000</pubDate><guid>https://digitaliziran.si/2026/10/26/the-half-life-of-an-instruction.html</guid><description>&lt;p&gt;Two French examples. That is all it took.&lt;/p&gt;
&lt;p&gt;I told DeepSeek&amp;rsquo;s larger model, V4 Pro, to translate every question into Russian. One earlier turn in which the assistant had answered in French was enough to make it slip 17 times in 100. Two, and it answered in French &lt;strong&gt;90 times out of 90&lt;/strong&gt;. Switch on thinking mode - the setting where the model writes out its reasoning before it answers - and the same model followed the instruction between 89 and 100 times in 100 at every number of turns I tested, up to 50.&lt;/p&gt;</description></item></channel></rss>