If our TVs were powerful enough to do this in real time without noticeable lag
The problematic source of Lag isn't processing overhead (that's actually entirely solveable even on weak hw by not processing in frames, but switching to something more immediate like scanlines), it's in the foundation of the algorithm itself. Interpolation is defined as having at least two data-points(frames) worth of history to interpolate between. Meaning you are automatically, even with infinite processing speed, at least doubling the input lag of the source.
Said AI technique is no exception - although I find it a bit amusing the speaker in the video talks about 'no artifacts' in the slow-motion video and then showing obvious occlusion artifacts right when he says it.
But I digress - if you want to do this without introducing meaningful lag you need to do motion-
extrapolation, like that Lucas Arts technique linked earlier in the thread. This is feasible, but it introduces its own sources of artifacts(in addition to artifacts interpolators have) - because now you have an algorithm make up(predict) new data. ML can help with this (it's the same problem set that DLSS and other temporal solutions are solving), and like all these temporal methods, it requires invasive integration (actually more invasive than DLSS), so no free-lunch with TV (or some other device/hw) doing it for all games.
Watch the arms on the astronaut in front. They completely disappear multiple times. Hilarious that the guy making the video chose that moment to say "I would have a hard time telling them apart from the AI generated ones". Maybe the ones where the arms are gone, that would be a good starting point.
Yea I think it's just the fact artifacts are smoothly introduced (there's little of the jarring pop, they sort of fade in and out of existence instead) that misled him to believe it's artifact free, but it has all the major problems of other motion interpolators.