Harshad Khadilkar

Over the past two years, Generative AI has been thrown at every problem, from its traditional forte of text and image generation to planning problems and time series forecasting. Keeping in mind that transformer and diffusion-based models are both probabilistic inference engines, the question in every practitioner's mind is: How can I leverage GenAI without leaving myself at risk if (when?) things go wrong? This talk will focus on the responsible use of GenAI techniques in decision-making systems. We will cover the strengths and weaknesses of generative approaches in quantitative domains and discuss ways their vast pre-training can be leveraged while retaining the robustness of automated systems. To help the discussion, we will discuss how (and more importantly, if) we can use LLMs in time series forecasting, logistics planning, and where diffusion models are a better bet than transformer-based architectures.