A good rule of thumb I always use if a science article title has the word "breakthrough"in it, then it's probably not a breakthrough.
If the title does nothing to describe the actual discovery made and solely consist of "Breakthrough in [Field]" then it's definitely not a breakthrough.
A good rule that could be refined by applying it only to topics that the general public cares about. A breakthrough in analytic number theory or international accounting standards is probably genuine, one in AI or battery technology probably not.
Why is that? I suppose complex maths is harder for the science journalist to understand, and doesn’t get as many clicks, so if they are reporting on it it’s because it’s substantial?
Any rule that reduces the posterior of a breakthrough is generally going to be an improvement. Unless your definition of "breakthrough" is extremely generous.
>"[its] training times are about 7-10 times faster, and... memory footprints are 2-4 times smaller" than those of previous large-scale deep learning techniques.
Which matches the abstract. If this has general applications it's a pretty big leap to shrink model sizes and speed up training by orders of magnitude, especially at a time when many SOTA models are only feasible for well funded groups because of their size.
This is not a strong result as noted by the reviewers, and they have not proven state-of-the-art performance among other things. Those of us who did ML in academia also look down upon using the media to bolster ones claims before a thorough peer review of research.
