‘Tokenmaxxing’ your way to a zero-day work week

IF YOU strain your ears enough, you might hear a faint, creaky squeaking on the periphery of your consciousness. That would be the sound of the hamster wheel speeding up at JPMorgan Chase. Either that, or you have tinnitus.

Tinnitus would be only marginally more discomfiting than what awaits white-collar workers as more firms start to peg artificial intelligence usage to employee performance.

In March, Business Insider reported that JPMorgan Chase has begun tracking its engineers through granular dashboards that categorise them as “heavy”, “light” or “non” users of AI tools such as GitHub Copilot. The engineers have ostensibly been given two core objectives – improve their coding performance and use AI to get more done.

To be clear, this isn’t a Luddite’s lament. If you’ve ever helped your boss install a printer, you know that there are those among us who view technological advances with the same suspicion that a cat reserves for a new vacuum cleaner.

So, employees need prodding at times, if only to ensure that they aren’t outpaced by the very tools designed to assist them.

But new technology cannot render obsolete the ancient quandary of Goodhart’s Law: When a measure becomes a target, it ceases to be a good measure.

As more companies turn to chatbot-flogging as a proxy for productive AI adoption, a plethora of unintended consequences will thrive as workers game a system that determines the size of their bonuses.

No company better illustrates this than Meta, an early adopter of AI-centric performance metrics that had reportedly gamified AI usage among its employees. At least you cannot say it doesn’t lead from the top – its CEO Mark Zuckerberg is said to be building an AI agent that will help him do his job.

Anyway, at one point, a Meta employee independently created a leaderboard that tracked how many tokens – a measure of how much data an AI model is processing – the company’s workers were using. This sort of leaderboard is a symptom of the “tokenmaxxing” fever endemic in Silicon Valley, where using more tokens conveys bragging rights and connotes productivity, however dubious that causal relationship.

Over 30 days, Meta employees on the dashboard collectively used more than 60 trillion tokens – the employee topping the league averaged 281 billion tokens. On the low end of assumed rates, this absolute legend alone would have cost Meta some US$1.4 million a month in token usage.

Meta’s chief technology officer Andrew Bosworth apparently has no problems with the expense, arguing that the productivity gains outstrip the cost. “It’s like, this is easy money,” Bosworth said. “Keep doing it. No limit.” My suspicion is that chief financial officers everywhere would beg to differ.

An existential consequence

Perhaps these firms will successfully incentivise genuine productivity. But then workers will run headlong into a more existential consequence as every incremental man-machine interaction makes the machine better at doing the man’s job.

The performance assessments will subject the employee to a peculiar, self-cannibalising logic, driving him to oil and prime the very guillotine that will eventually take his head off.

Just recently, JPMorgan’s CEO Jamie Dimon predicted that AI would cut the work week down to three-and-a-half days in 30 years’ time.

That sounds nice, until you realise that some workers could see that number slashed to zero days – either because they failed to use AI well enough, or because they used it far too well.