Amazon constructed an inner AI instrument staffed by 36 engineers. Its staff responded by gaming it.
The follow has earned its personal identify inside the corporate: “tokenmaxxing.” Staff are utilizing MeshClaw, Amazon’s internally developed platform for creating AI brokers to automate duties, not as a result of they want the automation, however as a result of they want the utilization numbers. The purpose is to look busy on AI leaderboards that monitor how continuously builders work together with the corporate’s instruments.
The leaderboard downside
Amazon mandates that over 80% of its builders use AI instruments on a weekly foundation. That utilization will get tracked and displayed on workforce leaderboards. Workers have began operating pointless automated duties by MeshClaw purely to inflate their token counts.
Amazon has stated these token metrics received’t issue into efficiency opinions. However employees report feeling anxious about low rankings no matter official assurances.
MeshClaw and the 36-engineer funding
MeshClaw was developed by a workforce of 36 engineers at Amazon. The instrument is designed to let builders construct AI brokers that automate varied duties.
The 80% weekly utilization mandate indicators to staff that there’s a minimal acceptable degree of engagement, no matter whether or not their specific workflow advantages from AI help.
A sample throughout Large Tech
Amazon isn’t alone on this. Comparable tokenmaxxing behaviors have been reported at Meta and Microsoft just lately. The sample is constant throughout the hyperscalers: make investments billions in AI infrastructure, mandate inner adoption, monitor utilization metrics, after which watch as staff optimize for the metric reasonably than the result.
Tech corporations have collectively poured monumental sums into AI applied sciences since late 2022, and they should present shareholders that the funding is paying off. Inner adoption charges are one of many best numbers to level to in a board presentation.
What this implies for buyers
For anybody watching the AI commerce, tokenmaxxing ought to increase a yellow flag about how inner AI adoption numbers get reported. If staff at Amazon, Meta, and Microsoft are inflating their utilization stats, then the adoption figures these corporations cite publicly could also be much less significant than they seem.
