<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>cycanic&apos;s workshop</title><description>AI ideas tested on real systems: measured results, code included.</description><link>https://cycanics.github.io/</link><item><title>[Part 1/2] A production LLM&apos;s weights edited almost-by-hand. It worked. A logistic regression did the same job better.</title><link>https://cycanics.github.io/articles/weights-vs-probe/</link><guid isPermaLink="true">https://cycanics.github.io/articles/weights-vs-probe/</guid><description>Weight-space programming tested on a 26B production LLM judge: a direct weight edit vs a linear probe vs a better prompt, measured.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate></item><item><title>[Part 2/2] Prompt optimization for an LLM judge: how to see how much accuracy your prompt leaves unused, and when to stop</title><link>https://cycanics.github.io/articles/prompt-optimization-gap/</link><guid isPermaLink="true">https://cycanics.github.io/articles/prompt-optimization-gap/</guid><description>A linear probe on the model&apos;s last hidden state shows how much accuracy an LLM judge&apos;s prompt leaves unused, and when prompt optimisation can stop.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate></item></channel></rss>