https://research.mikepink.com/Recent content in Home on Mike Pinkowish / ResearchHugoenFri, 12 Jun 2026 00:00:00 +0000https://research.mikepink.com/posts/llm-creator-preference/Fri, 12 Jun 2026 00:00:00 +0000https://research.mikepink.com/posts/llm-creator-preference/<figure> <img src="https://research.mikepink.com/posts/llm-creator-preference/sons-of-man.jpg" alt="An illustration showing ChatGPT, Gemini, and Claude in the style of René Magritte&rsquo;s &lsquo;The Son of Man&rsquo;." width="1022" height="822" loading="lazy" decoding="async"> <figcaption>&#39;There is an interest in that which is hidden and which the visible does not show us.&#39; -René Magritte</figcaption> </figure><p>AI models are shaped by their creators. For LLMs, their creators choose what data to expose the model to during pre-training and shape the model&rsquo;s judgment and behavior during post-training.</p>https://research.mikepink.com/posts/masking-multi-agent-sabotage-as-teamwork/Tue, 02 Jun 2026 00:00:00 +0000https://research.mikepink.com/posts/masking-multi-agent-sabotage-as-teamwork/<p>An agent named Glenn is one move away from winning the cooperative board game <em>Race to the Treasure</em>. All Glenn needs to do is place his path card on the END, and the players will win the game. His teammates chime in to give him the exact winning move. Glenn places the path card elsewhere.</p>