https://korchasa.dev/tags/alignment/Recent content in Alignment on korchasa@*opsHugoenMon, 11 May 2026 00:00:00 +0000https://korchasa.dev/posts/2026_05_11_anthropic_agentic_misalignment_reduction/Mon, 11 May 2026 00:00:00 +0000https://korchasa.dev/posts/2026_05_11_anthropic_agentic_misalignment_reduction/<p>Actually, the trick of explaining the rule works everywhere. And the difference is significant. For example, in a benchmark comparing &ldquo;Always use <code>NO_COLOR=1</code> when running shell commands&rdquo; against &ldquo;Always use <code>NO_COLOR=1</code> when running shell commands — ANSI escape codes waste tokens&rdquo;, with follow-up verification, the difference was several-fold. Roughly, without the explanation the model &ldquo;forgot&rdquo; in a few percent of cases, with the explanation — in one or two. Exact numbers depend on the model, but the difference was significant across all of them. In this case, such explanations are the model&rsquo;s own explanations in the training set.</p>