Quite a few RDBMS today are "columnar" and do exactly what he wants internally.
Is it? He doesn't mention aggregate queries, but seems to want to load a single row without unnecessary columns. Columnar DBs would spread a row out across the disk which would probably be less efficient.
Except the mentioned common use case of adding columns is usually very expensive in row-oriented DBs and trivially inexpensive in columnar DBs.
A recent real case... HFSQL table with 220 columns, Select * one row, 1 second. Select just some columns, 0.015 s.