Paul Graham argues that token counts are an inadequate measure for AI inference as models improve, advocating for a new metric focused on problem-solving capability. This suggests a potential shift in how AI services are valued and priced, impacting AI developers and users alike.
Paul Graham's call for a new AI inference metric highlights a growing disconnect between current pricing models (token counts) and the actual value delivered by increasingly sophisticated AI. This matters because it could fundamentally alter how AI services are bought and sold, shifting focus from raw computational output to problem-solving efficiency. AI developers like OpenAI and Anthropic, and their customers, are directly affected, as a new metric could lead to different pricing structures and competitive advantages. In the short term, this is a conceptual discussion, but long-term, it could drive innovation in AI model design and pricing. A key opportunity for traders lies in identifying companies that are best positioned to adapt to or even define these new valuation metrics, potentially gaining market share or improving profitability.