“Scaling purely autoregressive models by 50,000x from GPT-2 to GPT-4.5 (2019-2025) only improved performance on ArcOne from 0% to about 10%, and on ArcTwo from 0% to 0%.”