Their training, validation, and test set is split on time rather than on the person speaking so 'destroys' may be overly optimistic for real-world applications.
Is this all just better quality of pattern matching or there is more to it ? I'm guessing that only progress being made in AI is in scale of data crunching speed and size of data availability. Is there any research to move AI towards self awareness or let's say story writing ?
The immense and sudden success of deep learning seems to have recalibrated everyone's expectations. We're solving The Hard Problem here; of course people are working on it but you can't expect to overturn a whole field twice a year.
Deepmind have said they're working on memory and similar challenges, which are an important step to reaching the things you ask for.
No, lip reading was just an single ability of the machine. What distinguished it was, at least apparent, consciousness about itself and desire for self preservation.
