In this paper, we investigate the role of large language models in predicting innovation. We contrast two main paradigms: i) narrow experts: which consists of supervised and semi-supervised models trained or fine-tuned on a specific task and ii) competent generalists: which consists of large language models with zero-shot and few-shots learning. We define the task of innovation modeling and present the first attempt to understand the transformation from research to innovation. We focus on product innovation which can be defined as the process of transforming technology to a product or service and bring it to the market. Our extensive empirical evaluation shows that most existing pretrained models are not suited and perform poorly on the innovation modeling task. We also show that injecting research information helps improving the alignment from technology to the market. Finally, we propose a new methodology and fine-tuning strategies that achieve significant performance boosts over the baselines.