“Token Legend,” or: how Meta burned 73.7 trillion tokens and accidentally ran an experiment

Somewhere in Menlo Park, over the past few weeks, an AI agent hummed away for hours on end, researching things nobody would read, producing results nobody needed, while its owner had almost certainly gone home. The agent kept working regardless, because every wasted token brought its owner one step closer to the most coveted distinction Meta had to hand out internally: the title of “Token Legend.”

The numbers behind it are absurd enough that you have to read them twice. Meta's workforce consumed 73.7 trillion AI tokens within 30 days, which at list prices comes to about 221 million dollars, per month. An internal dashboard called “Claudeonomics” ranked the top 250 consumers among more than 85,000 employees and handed out honorary titles like “Token Legend,” “Cache Wizard,” and “Session Immortal.” The frontrunner alone reached 281 billion tokens. And according to SemiAnalysis, some employees left their agents grinding for hours on tasks with no outcome goal, purely to climb the ranking. When The Information reported on it, the leaderboard was scrapped within days, and CTO Andrew Bosworth left behind a line worth framing: “All motion is not progress and token usage alone is not a measure of impact of any kind.”

The reflex now is to laugh at Meta. I would suggest something more interesting: Meta has just handed us, involuntarily and under laboratory conditions, one of the most expensive behavioral-science experiments in business history, and we ought to take the results more seriously than the punchline.

The plan was flawless, and still did not work

First, in defense of the people who built the leaderboard, because their logic was impeccable. If AI usage raises productivity, and if tokens measure AI usage, then you should reward token consumption, and because people react to rankings the way moths react to light, you build a ranking. Every technocrat on earth would have signed off on this plan, it makes complete sense. Except it does not work, and in my line of work you learn early that not everything that makes sense works, and not everything that works makes sense.

What happened here has a name, and it is older than any language model. When the French colonial administration in Hanoi set out to fight a rat plague around 1902, it paid a bounty for every rat tail handed in, whereupon residents began breeding rats, cutting off their tails, and releasing the animals again so they would keep multiplying. The metric exploded, the problem grew. Goodhart's Law, as we now call the pattern, holds in Marilyn Strathern's formulation that once a measure becomes a target, it ceases to be a good measure. A token is a rat's tail with better PR.

So far, so familiar. The genuinely interesting question is a different one: why does this happen to Meta of all companies?

Why Meta of all companies

Because let us be clear, Meta is not a company that also happens to have something to do with behavior. Meta is arguably the most ambitious applied behaviorism project in human history, a corporation whose entire business model rests on shaping the behavior of billions of people with variable reward schedules, streaks, likes, and finely tuned notification rhythms. Skinner's pigeons pecked at levers, Meta's users swipe across glass, the principle is the same and Meta commands it better than any university. If any company on this planet ought to know how incentives work, it is this one.

And this very company builds itself an internal incentive system that a third-semester behavioral-economics student could take apart as an exam question.

Here is the point: that is not a contradiction, it is the solution to the riddle. Meta's entire expertise is in quantitative behavioral steering, more time on the platform, more clicks, more scroll depth, more sessions, and for that goal behaviorism is a wonderful tool, because the business model can be perfectly indifferent to which behavior is shown as long as the counter goes up. Whether someone scrolls out of joy, out of boredom, or out of compulsion is, for the advertising revenue, the same row in the same table. With its own employees, though, Meta suddenly needed something entirely different, namely qualitative behavioral change: better judgment, smarter use of tools, more considered code. And for that kind of behavior a different science applies, one that Silicon Valley evidently keeps nobody in house for.

The science has been clear for fifty years

That science has been remarkably clear for over fifty years. Edward Deci showed in 1971 that students paid to solve puzzles stopped playing sooner in the unpaid break than the unpaid comparison group, the reward had not reinforced their interest but replaced it. Lepper, Greene, and Nisbett repeated it in 1973 with nursery-school children and felt-tip pens: those who received an announced certificate for their drawing subsequently drew markedly less in free play than the children who had simply drawn. And Sam Glucksberg had demonstrated as early as 1962 that monetary incentives made people slower rather than faster at the candle problem, a task that requires a small creative detour in thinking, because reward narrows the field of view, which helps on simple diligence tasks and harms anything that demands thought. Extrinsic incentives work splendidly as long as the target behavior is simple, repetitive, and countable, they turn ineffective the moment quality enters, and they turn toxic the moment you need judgment.

The one test that separates the label from the behavior

In my work I call systems like the Meta leaderboard bribery design, and there is a simple falsification test with which any board can check any incentive program in five minutes: take the reward away and see what is left of the behavior. If the behavior dies, then you never changed behavior, you merely rented activity, and rent, as we know, runs only as long as you keep paying. The elegant thing about the Meta case is that the company has just run this test in public for us: leaderboard switched off, and the tokenmaxxing evaporates, the agents idling for hours disappear, because the tokens were never the point, the rank was. A behavior produced exclusively with a ranking can be erased entirely with that same ranking, and that is precisely how you recognize that nothing was learned, nothing internalized, and nothing changed.

Behavioral capital, the asset class with no line on the balance sheet

One could take all of this for a Silicon Valley anecdote, were it not for the uncomfortable arithmetic behind it. Companies have their finances audited to the cent, their supply chains certified, and their software penetration-tested, but the behavior of their people, the raw material from which performance is produced in the first place, they shape by gut feeling and with tools whose ineffectiveness has been documented since the seventies. For this asset there is not even a line on the balance sheet, though behavioral capital would be the most honest name for it, and what Meta did with it would, in any other asset class, be called negligence. If even the company with the greatest density of behavioral data and behaviorists in the world reproduces a textbook classic on its first internal attempt at qualitative behavioral change, then it is worth asking what is happening right now in the AI rollouts of the companies that do not even have a dashboard for it.

Bosworth is right, of course, motion is not progress. But the sentence came after the bill for 73.7 trillion tokens, and the cheaper order would have been the other way around.