The imagenet results’ “main” contribution was showing that scaling the compute on large datasets worked surprisingly well. They didn’t contribute any new models

The imagenet results’ “main” contribution was showing that scaling the compute on large datasets worked surprisingly well. They didn’t contribute any new models as much as they showed that scaling is what helped these models learn and obtain SOTA. It’s no surprise that industry has been pushing for more compute ever since.