This concept of making the best use of any available data to infer useful probabilities seems like a crude form of some ideas behind current LLMs, BUT they (Peters, Graham, Robinson and Woodhead) were doing this work 20 years earlier.