I agree with 95% of what he says, but I'd make one change to the Unix philosophy: text files aren't the best primitive.

Text files have a super low-level API (stat, seek, read). I think the system he describes would be much more usable if he picked slightly higher-level data abstractions. I would recommend Python's built-in types: list, tuple, dict, set. Most data can be easily represented by these types and they provide a higher level API than streams (files), which means the barriers to using them are much lower.