20 years ago that was true. MNIST and CIFAR-10 were a vehicle to show that your method had promise.

Today, any bad idea you throw at a wall will work on MNIST and CIFAR-10. Even buggy broken code will generally work. We've advanced to the point where these datasets are totally meaningless.

In 1910 if you could show that your airplane idea could fly 100ft, you had something promising. In 2010 we're so used to building airplanes and understand the physics involved to the point where this isn't a meaningful test of the promise of your new aircraft concept.

It's worse than "in mice" in biology/cancer papers. We cure everything in mice. Basically nothing ever transfers to humans. Same with MNIST/CIFAR-10. Everything works, basically nothing matters.

Or another way to put it. A "Hello World"-level compiler would be interesting to report on in 1960. That same compiler would be trivial today and anyone could build it in minutes/hours.