Are you basically saying to make pixels by averaging the color of the adjacent ones?
No, he isn't.
If its surprising to you that its possible to do better prediction of the missing information between the pixels of images than various kinds of interpolations (such as bilinear (aka "averaging" for 2x), bicubic, etc..), then its because you are wholly ignorant on the subject. Amazing that you jumped in to discuss a subject that you are wholly ignorant on.
If it were purely random data then your from-ignorance consideration of the subject would have some weight, but it isn't arbitrary data. Nobody spends their time looking at images of random data. The images we capture with photographic equipment have high degrees of coherence between neighboring scales (because thats the way the universe is,) and that necessarily includes the scales just above the Nyquist frequency.
Nobody is suggesting red dwarfs "enhance!" is possible, but a bit of upscaling can be done a lot better than simple interpolations.