I think you misunderstood. The authors created the very issue they are "debunking." They took GPT3, slapped on some random metrics and showed that these metrics

I think you misunderstood. The authors created the very issue they are "debunking." They took GPT3, slapped on some random metrics and showed that these metrics don't show scaling behaviour correctly, while the original publication of GPT3 actually did it correctly in the first place.