Michele Catasta, Replit’s head of AI, criticized internal company leaderboards where employees compete based on the amount of AI tokens they consume. Catasta said this leaderboard has turned into a display of showmanship rather than productivity, creating a "dystopian" image.
Speaking at Web Summit Rio, Catasta explained that in some companies, employees are encouraged to spend more tokens just to appear higher on the leaderboard. According to him, this competition does not reflect whether AI usage is actually generating value.
Token Count Does Not Reflect Performance
Catasta emphasized that token consumption is not a suitable indicator for measuring employee performance. He stated that using more tokens does not necessarily mean having a greater impact within the company.
He also noted that rewarding AI usage solely based on quantity can encourage unnecessary spending. Catasta likened excessive use to leaving the lights on at home without caring about the electricity bill.
Growing Debate Among Companies
This statement comes at a time when discussions among company leaders about the best way to measure AI usage are intensifying. Many executives are reconsidering whether increased usage automatically translates to higher productivity.
Last month, Amazon announced it had removed its internal AI usage leaderboard. A company spokesperson said the initiative was not designed to promote AI use just for the sake of usage.
During the same period, Uber’s Chief Operating Officer Andrew Macdonald said he did not see proportional productivity gains despite rising AI costs.
Charles Holive, Chief AI Officer at BNP Paribas CIB, described this approach, known as "tokenmaxxing", as a "vanity metric".
While Catasta expressed satisfaction with the recent shift in narrative around tokenmaxxing, he pointed out that this habit increases energy consumption and reduces capacity that could be allocated to other uses.
He conveyed the message that the focus should be on how correctly and efficiently AI is used, rather than how much it is used.
