What is it? Incentivizing people to use as many AI tokens as they possibly can. Some companies are even creating leaderboards to track who is using the most. Many companies report getting far more AI incidents, bugs, etc. But some people swear by it, saying it’s making them more forward with AI and getting their employees to dig into these tools more
There are both good and bad incentives to tokenmaxxing. Any time you start measuring something and rewarding people who look good on that measurement, you will get people gaming the system, and in this case, using tons of tokens but not building anything valuable. But you will also incentivize people to push themselves to explore and improve at things they wouldn’t have otherwise.
Is this practice of tokenmaxxing resulting in tons more AI slop? Absolutely. Is it also getting people who aren’t diving into AI tools to use them more and become better at agentic engineering? Absolutely.
My opinion: good thing if done right. Everyone should be figuring out how to use all the tokens they can while also making them useful.
So, I think companies and managers should be careful how they incentivize usage, but I still think they should incentivize it. There’s no perfect balance, but it’s better to try and get it wrong, and possibly waste some money and introduce a few more bugs, than to sit around and hope that employees just figure it out.
Links#
What it is#
- The Pulse: “Tokenmaxxing” as a weird new trend — Pragmatic Engineer overview
- What Is Tokenmaxxing? — Built In
- What Is ‘Tokenmaxxing’? — Inc.
- The Era of Tokenmaxxing
Leaderboards in the wild#
- Meta’s “Claudeonomics” dashboard — Fortune — 60T tokens in 30 days across 85k employees, then shut down
- Meta shutters internal AI token leaderboard — The Information
- Reid Hoffman weighs in on tokenmaxxing — TechCrunch
- Techies debate the leaderboards — AOL
- Token consumption: productivity metric or vanity trap? — Trending Topics
The bugs / incidents side#
- AI code creates 1.7x more issues — CodeRabbit
- Incidents per PR up 23.5% — QASource
- Are bugs and incidents inevitable with AI coding agents? — Stack Overflow
- AI-authored code needs more attention, contains worse bugs — The Register
- 43% of AI-generated code changes need debugging in production — VentureBeat