Hacker News Reader

> And in the case of AI, if I can’t train the model from source materials with public source code and end up with the same weights, then it’s not open.

Making training exactly reproducible locks off a lot of optimizations, you are practically not going to get bit-for-bit reproducibility for nontrivial models

View on HN

Nobody's asking for exact reproducibility — if the source code produces the software and it's appropriately licensed then it's Open Source.

Similarly, if you run the scripts and it produces the model then it's Open Source that happens to be AI.

To quote Bruce Perens (definition author): the training data IS the source code. Not a perfect analogy but better than a recipe calling for unicorn horns (e.g., FB/IG social graphs) and other toxic candy (e.g., NYT articles that will get users sued).

That's kind of true for normal programs as well, depending on the compiler you use, and if it has non-deterministic processes in it's compilation. But still, it's about being able to reproduce the same build process, and get a true realization of the program, even if not bit-for-bit, it's the same intended program.

Reply on news.ycombinator.com