It's the thing with abstractions: changing angle changes the abstraction by seemingly replacement candidates, but only the good ones maps properly onto physical hardware and human cognition.
eg: "everything is a file" maps properly to both physical storage with addressable, streamable content and to the human mind as data reachable by classification (file name). Maybe a revolutionary OS design could better that with "everything is a message", but I think you can understand why one can have doubt on this translation from files.
The difference between a file and a message is that a file does not prescribe format of the file - it's usually application dependent, but is understood to be plain text.
A "message" tends to have formats associated with it (which is what differentiate it from being just a "file" that's transferred via a protocol). These formats, like JSON, or xml, or whatever new fangled formats kids these days use, now requires architectural astronauting; namely, common field names, or standards, so that different programs would parse them similarly. And now you'd want schemas for those formats, and automatic parser generated for those formats, and more and more...
eg: "everything is a file" maps properly to both physical storage with addressable, streamable content and to the human mind as data reachable by classification (file name). Maybe a revolutionary OS design could better that with "everything is a message", but I think you can understand why one can have doubt on this translation from files.