AI Scaling vs. Human Scaling (www.tobyord.com)

Recent discussions in the AI community have raised important questions about how AI models scale with more tokens in comparison to human performance when given more time. This topic explores whether expanding an AI's token budget—akin to increasing the time available for humans—yields equivalent improvements in performance. Initial data suggests that while humans often gain significantly more from time allocated, AI performance improvements are less proportionate, necessitating up to 10,000 times more tokens for similar gains. As new models emerge, there's evidence of improving efficiency, indicating that the gap between AI and human scaling may be closing. This exploration is significant as it impacts our understanding of AI capabilities and their potential for self-improvement. The scaling dynamics could influence predictions about AI's recursive self-improvement (RSI) and its overall ability to reach superintelligent levels. The findings challenge the long-standing analogy between AI token scaling and human time scaling, suggesting that while AIs are improving, they currently still scale less effectively than humans. Ongoing research into these scaling relationships will be crucial for comprehensively understanding AI's trajectory and its implications for future intelligence development.