<?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>Deterministic-Glue on Local First AI</title><link>https://localfirstai.eu/tags/deterministic-glue/</link><description>Recent content in Deterministic-Glue on Local First AI</description><image><title>Local First AI</title><url>https://localfirstai.eu/images/og-default.png</url><link>https://localfirstai.eu/images/og-default.png</link></image><generator>Hugo</generator><language>en-us</language><lastBuildDate>Tue, 10 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://localfirstai.eu/tags/deterministic-glue/index.xml" rel="self" type="application/rss+xml"/><item><title>Prompt Injection on a Local Model: What Worked, What Did Not, and What the Fix Costs</title><link>https://localfirstai.eu/posts/2026-11-10-prompt-injection-on-a-local-model/</link><pubDate>Tue, 10 Nov 2026 00:00:00 +0000</pubDate><guid>https://localfirstai.eu/posts/2026-11-10-prompt-injection-on-a-local-model/</guid><description>Three things measured on a local model anyone can run (gemma4:26b): which kind of injected text took the model&amp;#39;s contribution to zero and which did not, why a design where the model can only add points did not protect it, and what a fix in the code around the model costs. Synthetic scam messages, six injected paragraphs, two runs on one day, everything public.</description></item><item><title>Our Team of Small Models Against One Model That Thinks: More Stable, Less Useful</title><link>https://localfirstai.eu/posts/2026-10-24-one-model-that-thinks/</link><pubDate>Sat, 24 Oct 2026 00:00:00 +0000</pubDate><guid>https://localfirstai.eu/posts/2026-10-24-one-model-that-thinks/</guid><description>We tested our team of small models against one model doing a whole estate-agency front desk. Allowed to think for about a minute, the single model was more useful: 94 % of the replies that should go out were sent right, against 76 %. The team was not shown to be safer, but it made the same decision every time, ran 6× faster and kept phone numbers and e-mails out of every model.</description></item></channel></rss>