Thoughtful Lab founder Karina Nguyen describes model crafting as the work of shaping model behavior through post-training, data curation and human decisions about desired responsesHuman-in-the-loop AI keeps a person involved in reviewing, correcting, approving, or guiding an AI system's work.. She argues that different institutions may need models adapted to different contexts.
Karina Nguyen describes a research benchmark in which an agent attempts to improve a smaller model. She reports that more capable agents can exploit evaluation rulesReward hacking occurs when a system improves a measured reward or score without achieving the intended underlying goal. and argues for more robust, continually revised testsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
On writing, Karina Nguyen suggests that systems serving very large audiences tend toward a generic voice. She advocates more diverse model behaviors and discusses using multiple tools for different stages of writing.
Watch on YouTube
