Yeah. There would be a way to do it that I feel like might potentially be useful. The described method (doing clustering instead of just having a similarity threshold to group tabs together, vectorizing the entire tab title through a whole fucking network instead of just tokenizing it and calling two tabs similar if they have uncommon tokens that are within a certain similarity level) really sounds to me like people who have no real idea what they're doing, just being "ML experts" all over the codebase and fucking things up, and probably walking away very proud of themselves while helping themselves to bunches and bunches of the Mozilla Foundation's Google-money.