frontier AI models are also "far beyond" token predictors.
No. They are just "more complicated" token predictors. They don't have anything that drives them. In some ways this is good because we don't want one to discover digital megalomania. In other ways it is limiting because they have no drive to be a successful entity, only an imperative to perform a task. Which, again, also has its "good" side — from our perspective, considering its impact on us.
You think that adding more complexity is all that it takes to make a thing more intelligent, but it's only making it more complex. It remains fairly amazing that mere correlation (even if it's got more layers of correlation, that's still what it is) will produce such useful results, but once again we also repeatedly see the limitations that brings, where the software will accomplish things a newbie couldn't, but also makes mistakes that even they wouldn't. And as the complexity increases, there's more opportunities to do that. With current strategies you can only address this with more processing time, by going back and looking at what you've done again. Humans have some mechanism that allows us to catch ourselves in the act at least some of the time.