Amazon recently shut down an internal leaderboard called "Kirorank" which tracked employees' use of its Kiro developer platform and scored them based on AI activity. The company, through senior vice-president Dave Treadwell, acknowledged that while the leaderboard was created with good intentions to raise awareness for AI's potential, it led to employees "tokenmaxxing" or artificially inflating their consumption of AI tokens, thereby increasing Amazon's computing costs. Treadwell explicitly instructed staff to prioritize building better products over focusing on token usage.
This move comes as Amazon faces significant capital expenditures, with an expected $200 billion spending this year, largely directed towards AI and data center infrastructure. The company also uses AI models from Anthropic, which recently switched to a consumption-based pricing model, further increasing costs for large customers like Amazon. The incident highlights a broader challenge for tech companies: how to encourage AI adoption without inadvertently incentivizing wasteful behavior and driving up infrastructure expenses, as seen with Meta employees similarly trying to boost their internal scores.
Amazon confirmed the "beta dashboard was not a formal or approved tool, and has since been deprecated." The company is now shifting its focus to measuring the success of its AI tools and the adoption of the technology through "normalized deployments," which indicate engineers consistently using AI to create valuable code, rather than simply tracking token consumption. This strategic adjustment aims to align AI usage with productive outcomes and operational efficiency.
