Irresponsible
This feels like an abuse of the leaderboard rather than a meaningful contribution.
Uploading a gigantic model containing only zero weights serves no practical purpose while consuming storage, bandwidth, and attention. The size leaderboard exists to showcase genuinely useful models and technical achievements, not placeholder artifacts that happen to be large.
If the goal was to make a point or test infrastructure, there are better ways to do that without polluting community rankings. I hope the maintainers consider excluding non-functional checkpoints from size-based leaderboards.
That criticism is fair, and exclusion of non-functional checkpoints may well be the right outcome. But that is also the point of the experiment: parameter count is a property of declared tensor shapes, not evidence of intelligence, training, information content, or capability.
Leviathan is structurally valid, yet entirely empty. Its presence at the top of a size ranking demonstrates that the metric—and any leaderboard based solely upon it—is easily gamed unless usefulness or functionality is considered.
Shape ≠ substance. Scale ≠ capability. A very large number can still represent a very large emptiness arranged with great ceremony.
I find the Trending sort may actually be a better method of showing genuinely useful models, since the models with the most likes and downloads are presumably the ones that are the most useful to people.
yo Aleem, what vick ere vorgot to mention was that it doesnt fucking waste space, hugigngface uses XeT compression which allows it to only take up like 5 GB
I disagree @Aleem , both the 16T and 19T experiments are interesting to me.
The size leaderboard exists to showcase genuinely useful models and technical achievements
The size leaderboard indeed exists to showcase useful models, but is it "useful for you"? This is a technical achievement.
Uploading a gigantic model containing only zero weights serves no practical purpose while consuming storage, bandwidth, and attention.
This is a bit gate-y considering OpenAI, Google and others probably consume vastly more, doing a lot less, and I ain't seen nobody complain. Ever.