Worth noting that Jeopardy Watson had very little ML and absolutely no deep learning (it was a few years pre-AlexNet). I don’t even think it used any neural networks; certainly not in any major way because they’re not a major topic anywhere in the press releases, working group notes, or the papers published by the Watson research group. Watson was an incredibly complex mixture of bespoke implementations of “classical” AI and NLP techniques to handle questions of different classes by transforming them into search & information retrieval problems. They were able to make it work pretty well for the very limited domain of questions that arise in Jeopardy, but it was also obviously a Herculean task to generalize that approach. I can totally believe that as executives started to grok what Watson really was they realized that it had more value as a brand than as a technology.