Reproducibility seems to be headed towards irrelevance in ML research. Is it too late? [D]

I feel that reproducibility is now a lost cause in machine learning research for three reasons:

- Many research is moving towards the physical AI territory, where you need expensive hardwares or even entire laboratories with high-speed cameras, in order to perform an experiment. You truly have no idea if the experiment can be reproduced and have to trust the demo. But demos are not perfectly reliable. Plus people are incentivized to only show the part of the demo that works. The entire system can fall apart the moment the recording stops.

- You have big AI companies releasing various tools, which they claim to solve a host of problems with certain amount of accuracy or efficiency. Unless you work at those companies there is really no proof of that and you will have to take their words on it. They have strong financial incentive to blow-up those figures. There is no solid way to check it either because the problem that they solve are so vague and subjective.

- We need to address the elephant in the room which is that people are incentivized to produce non-reproducible work to prevent their lunch being eaten by their competitors or looking bad. That's why some of us will probably never get a reply when we email the authors for their code.

So what now? Maybe everything will be OK because we can contrast it with scientific progress in earlier parts of history, e.g., building the atomic bomb or sending people to the moon. These projects had low "outside reproducibility" but high "internal reproducibility". Plus all these work were mathematical in nature and carefully checked. But I don't think many areas of machine learning research is like that. What do you think? Should reproducibility be abandoned? If not how is it best implemented going forward?