Tag
New research indicates that the apparent 'global convergence' in scaled AI models is actually a statistical illusion caused by selection bias in model width and depth, and disappears once calibrated.