[Cognition CEO Suggests Replacing Token Usage Metrics with Actual Output Evaluation] According to monitoring by Beating, Scott Wu, co-founder and CEO of AI programming startup Cognition, stated on the Founders podcast that some companies have gone "too far" by assessing programmer performance based on the number of AI tokens consumed by employees. He suggested that actual completed tasks and output should be the standard for evaluation. This trend, known as tokenmaxxing in Silicon Valley, refers to employees reflexively overusing AI programming tools like Claude, Codex, or Cursor to gain an advantage on internal AI usage dashboards or performance reviews. Jacob Lauritzen, CTO of legal AI startup Legora, similarly criticized this practice on a podcast this month, noting that showcasing token usage in performance evaluations leads employees to inflate usage purely for better-looking data, calling it a foolish evaluation method. Andrew Feldman, CEO of chip developer Cerebras Systems, remarked at this month's Bloomberg conference that providing employees with unlimited token allowances was "extremely foolish" from the start. Feldman advised companies to focus more on cost-effectiveness when using AI tools, opting for low-cost open-source models rather than overusing top-tier closed-source models for routine tasks. [Original Link]