“Despite a 50,000x increase in scale, LLM accuracy on the ARC-1 benchmark only improved from 0% to approximately 10%.”