It was always the data, not the code. Try and find an open dataset for any interesting machine learning problem and you'll realize that while "freedom" was busy doing things like setting back the use of precompiled headers in GCC a decade and making it virtually impossible for an artist to get a copy of ffmpeg that handles all the file formats she needs, the real value remains the data.
We don't need 15 open source PDF viewers. We need open access to the papers. And even hippie scientists at Berkeley seem unwilling to share those for some reason. So odd given the heritage.
I don't care much about Google's half-baked machine learning library. Give me the 128k neural output from the 250TB of voice queries if you wanna be "open" and advance machine learning. Unsurprisingly they've got that locked up tight. But culturally you can make the argument that's very much "ours" just like government-funded research papers are.
Given interesting data, nerds will ALWAYS find a way to read it. Focusing on code was a bit of a mistake; that's cheap and you get it for free. And the gap between open software licenses and Creative Commons licensing always seemed odd.