The Data Factory: Inside the $100B Race for Post-Training Supremacy, with Labelbox CEO Manu Sharma

Frontier AI development has shifted from pre-training toward post-training, where models are refined through reinforcement learning to master complex, long-horizon tasks. This evolution necessitates a transition from basic data labeling to the creation of sophisticated environments where models develop skills like coding and mathematical reasoning through automated feedback and verifiable rewards. While traditional fine-tuning remains relevant for specific efficiency goals, many enterprises are increasingly relying on context engineering to leverage the robust reasoning capabilities of existing foundation models. Manu Sharma, CEO of Labelbox, highlights that top-tier AI labs now invest over a billion dollars annually in training data, utilizing a global network of domain experts to build these specialized reinforcement learning environments. This data-centric paradigm underscores the critical role of high-quality, expert-driven data in achieving reliable, superhuman performance across diverse professional domains.