Oh How Things Change
One of the annoying things about AI is that sometimes it provides invaluable assistance that materially improves output, but sometimes it is an enormous distraction and a waste of time. This is a consequence of what’s known as “the jagged frontier” of AI. (h/t Ethan Mollick et al.)
Tokenmaxxing was the job
The problem is you never know which you’ll get, and because there’s a chance you get the former, it’s almost always worth trying. This was the fundamental principle beneath the idea of tokenmaxxing, and not two months ago, an employee at several elite tech companies may have been put on notice if they weren’t tokenmaxxing. Interacting as much as possible with the AI agent of choice was a (mistaken) pre-requisite for doing your job well.
The flex is now spending less
As with writing, however, it is becoming increasingly high-status to REDUCE your token spend.
But you only begin to understand which problems AI is *actually* helpful for, however, when you use AI enough to learn where it sucks. So hurry up and get distracted by AI!
THAT may actually be a critical input to ultimately learning how to use less of it.
FAQ
What is the "jagged frontier" of AI?
The uneven edge of AI capability (h/t Ethan Mollick): the same model that materially improves one task wastes your afternoon on a neighboring one, with no reliable way to predict which side of the line you're on except experience.
Is tokenmaxxing still a good idea?
It had its moment — two months ago heavy agent use was table stakes at elite tech companies, and skipping it could put you on notice. Now the flex is inverting: use AI hard enough to learn where it sucks, then cut your token spend and keep the wins.
Should teams be using more AI or less?
More first, then less. Heavy use is the only way to learn where AI actually earns its keep, and that knowledge is what eventually lets you spend fewer tokens on better problems.