Nathan Lambert of AI2 announced a podcast and lecture with Scott Geng examining the practical realities of post-training the Olmo 3 model via direct preference optimization. The conversation addresses what is required for a research idea to reach a near-frontier model, the frequently messy data aspects of DPO, and organizational sequencing challenges that arise when multiple teams collaborate on open models. The session is presented as a rare public case study of these post-training workflows.