Just a few weeks ago I was contacted by Amazon recruiter and refused for exactly this reason. I expect more layoffs as they figure out they don't need this many engineers after all. They will turn into money pumping google search analog.
It's all probabilistic, my guess. I.e. model produces probabilities for a set of actions from the same video. Even pretended action may look more like it than anything else. Thus getting higher probability.
"conceptual framework" can actually be another generalist model. Splitting model also comes with some advantages. Like easy separate tuning and replacements. Easy scaling by simply duplicating heavily used model on new hardware.