Eric Siu explains token maxing: spending excessive compute on AI without measuring results. He warns this creates perverse incentives and recommends pairing token usage with customer impact or outcomes to prevent wasted costs and ensure AI experiments actually deliver value.
Here's the uncomfortable truth about token maxing. What token maxing is, first and foremost, is spending as much as you can on tokens. If you work for me, you have to spend as much as possible. I would just run my tokens on stupid tasks all day. I may just have an agent just to hit that number so I don't get fired. Just token maxing alone is pretty stupid because it creates perverse incentives. When you create a metric like this, you have to have a pairing metric. A pairing metric means, sure, you can try to token max, But at the end of every month, you have to go to your manager and say, hey, here's all the things that we built, here's how it helps customers, or here's how it gets us customers. Because if you don't have a pairing metric there, it's going to just drive a bunch of compute up. You're going to pay an arm and a leg, and, uh, you're going to get nothing from it.
