https://moehlrt.com/blog/Recent content in Blog on Moritz EhlertHugoen-USCopyright © 2026, Moritz Ehlert.Tue, 14 Apr 2026 00:00:00 +0000https://moehlrt.com/blog/open-character-training-does-model-size-matter/Tue, 14 Apr 2026 00:00:00 +0000https://moehlrt.com/blog/open-character-training-does-model-size-matter/<style> /* Mobile Layout Fixes (Screens smaller than 768px) */ @media (max-width: 768px) { .responsive-tldr { flex-direction: column !important; margin-left: 0 !important; } .responsive-tldr-label { width: auto !important; margin-bottom: 0.5em !important; font-weight: bold; } /* Bilder-Fix für mobile Größen */ .responsive-img-wrapper { margin-left: 0 !important; height: auto !important; width: 100% !important; overflow-x: auto !important; overflow-y: hidden !important; /* Versteckt den abgeschnittenen Bereich */ -webkit-overflow-scrolling: touch; padding-bottom: 10px; } .responsive-img-wrapper img { width: 145% !important; min-width: 550px !important; max-width: none !important; /* --- HIER WIRD ABGESCHNITTEN UND VERSCHOBEN --- */ /* Mache die Minus-Werte größer (z.B. -50px), um mehr abzuschneiden. Für margin-left: Ein positiver Wert (z.B. 20px) schiebt das Bild nach rechts, ein negativer Wert (z.B. -30px) schneidet links ab. */ margin-top: -35px !important; /* Oben abschneiden */ margin-left: -57px !important; /* Links abschneiden */ margin-right: -30px !important; /* Rechts abschneiden */ } /* --- HIER WIRD UNTEN ABGESCHNITTEN --- */ .responsive-img-wrapper.phone-crop-img-1 img { margin-bottom: -475px !important; /* Bild 1 unten abschneiden (Hier den Wert anpassen!) */ } .responsive-img-wrapper.phone-crop-img-2 img { margin-bottom: -150px !important; /* Bild 2 (Mitte) unten abschneiden */ } .responsive-img-wrapper.phone-crop-img-3 img { margin-bottom: -300px !important; /* Bild 3 (Unten) unten abschneiden */ } .mobile-margin-reset { margin-left: 0 !important; } } </style> <h1 id="tldr" class="post-heading"><a class="post-heading-anchor" href="#tldr" aria-label="Link to this section">#</a>TL;DR</h1> <p>I replicated and extended character training for LLMs—the technique used by Anthropic, OpenAI and other frontier AI labs to shape AI assistant personas. Starting from the <span class="post-link-wrap"><a href="https://arxiv.org/pdf/2511.01689" target="_blank" rel="noopener">Open Character Training</a><span class="post-preview post-preview--url" role="tooltip">https://arxiv.org/pdf/2511.01689</span></span> paper, I implemented the full pipeline using the <span class="post-link-wrap"><a href="https://thinkingmachines.ai/tinker/" target="_blank" rel="noopener">Tinker SDK</a><span class="post-preview post-preview--og" role="tooltip"><span class="post-preview-og-img" style="background-image:url('https://thinkingmachines.ai/news/announcing-tinker/images/tinker-cover-social.png')"></span><span class="post-preview-og-body"><span class="post-preview-og-site">Thinking Machines Lab</span><span class="post-preview-title">Tinker</span><span class="post-preview-og-desc">Tinker is a training API for researchers and developers.</span></span></span></span>, trained multiple personas with a focus on sycophancy due to its severe implications for today&rsquo;s SOTA models. I also went beyond the original work by conducting RLAIF comparisons and scaling experiments across model sizes (8B vs. 70B), exploring the question: does model size matter when shaping an assistant&rsquo;s persona?</p>