I think this is a bit optimistic compared to my view (wrt portability). There's a large stack of software that is involved in training and probably less so in inference. I'm not saying it's impossible but there are definitely different levels of reproducibility and the academic incentive structure doesn't really prioritize reproducibility in my experience. I'm sure it varies quite a bit, I'd be curious to know how those in this problem space are thinking about reproducibility and at what level.

I know it sounds unrealistic and not aligned with academic incentive structures. But those are the exact structures that gave us a lot of headaches in the experimental sciences. I think it would be a good north star to aim for something that resembles how those are trying to address the reproducibility crisis. Better than to embrace the most black-box version of math that AI systems can produce (million-line proofs without context). Even if a reproducibility crisis is seemingly impossible (although agents so far have also been pretty good at finding compiler bugs).