Today in Shanghai, it’s about 22 degrees, with intermittent clouds in the sky.
I originally planned to go out for a stroll, but these cloudy days always leave me feeling a bit sluggish, so I decided to brew a cup of tea and browse through some tech gossip at home.
As a result, I stumbled upon something extremely absurd.
The Crazy Metric Game
There’s a highly ridiculous, perverse trend sweeping through the industry right now called “Tokenmaxxing.”
People generally translate it as maximizing token usage.
Sounds pretty geeky, right?
But here’s the problem.
This stuff isn’t being used for any groundbreaking innovation at all; it has purely devolved into a crazy, grassroots metric-chasing game ╮(╯▽╰)╭.
The origin sounds incredibly cliché.
Bosses grew envious of the latest trends in Silicon Valley, slapped their foreheads, and issued a strict mandate—every employee must embrace large language models.
How to implement this?
The executives came up with what they thought was a brilliant idea: track how many AI tokens each employee consumes daily.
Using more means you’re highly motivated; using less means you get reprimanded.
Honestly.
Anyone with half a brain could guess the outcome of this forced quota.
(This screen looks advanced, but in reality, everyone is just secretly farming points at the bottom)
When Compute Becomes “Cyber Exhaust”
The developers at the bottom were driven crazy and immediately started frantically inflating their metrics.
A major ride-hailing giant (let’s say, one as familiar as Uber) became a massive joke internally.
It was truly wild.
They originally approved a large sum of funding, intending to reserve it as the AI compute budget for the entire year of 2026.
Guess what?
In just three short months, this money was completely burned to ashes by the employees.
Everyone frantically bombarded the LLMs with redundant nonsense and meaningless automation scripts, all just to score high on internal performance reviews.
This tactic looks exactly like how we used to leave our phones downloading large files overnight just to use up expiring data plans at the end of the month.
To give you a sense of how surreal this involution is, let’s look at what Meta was doing.
It wasn’t that simple.
Meta set up an internal, company-wide leaderboard.
Tens of thousands of employees competed fiercely on the chart, vying to see who could consume the most tokens.
Everyone tried to outdo each other, pushing the monthly consumption to the staggering level of tens of trillions.
Crazy, right?
Even Mark Zuckerberg himself couldn’t squeeze into the top 250.
Old employee Xiao Wang might just need to send a simple leave-of-absence email.
But he insists on having the AI rewrite it twenty times in Shakespearean style, wasteland punk style, and Victorian English.
Code veteran Lao Li is even more ruthless.
To top the leaderboard, he simply set up dozens of AI agents in the background, letting these AIs endlessly chatter nonsense with each other in a pointless loop.
How could what they’re burning be considered genuine demand?
It has all turned into a pile of nutrient-free cyber exhaust.
Speaking of which.
Many of the models used by these tech giants come from top-tier companies like Anthropic.
Previously, Anthropic offered very generous subsidy plans for enterprise tiers, somewhat like an all-you-can-eat buffet.
But watching these folks come in every day with massive troughs, eating some and dumping the rest, Anthropic finally couldn’t stand it anymore.
They publicly stepped out and fired shots.
They truly couldn’t hold back.
They bluntly exposed this farce, stating that such behavior is artificially creating a massive bubble of inflated compute demand.
Immediately after, Anthropic decisively canceled those bottomless enterprise subsidies.
Flipping the table like this was indeed a beautiful move.
Clean and crisp.
You take the compute I provide at a loss to fool your own bosses, do you really take me for a sucker?
This slap in the face directly tore off the fig leaf of the industry’s false prosperity.
Those Heavy “Cyber Scrap Nails”
Speaking of this, it reminds me of something else.
History often shares striking similarities.
There is a classic concept in economics called “Goodhart’s Law.”
The general idea can be summarized as: once a measure becomes a target, it ceases to be a good measure.
A few days ago, I was chatting with netizens in a group about the funniest anecdote related to this. (takes a sip of tea)
It took place during the Soviet planned economy era.
At the time, authorities issued strict production quotas to a nail factory.
Initially, the leadership calculated performance based on “the number of nails produced.”
Upon receiving the quota, the workers turned around and started modifying the assembly line.
They worked day and night, producing countless micro-nails as thin as pins.
The quantity was definitely over-fulfilled.
But what they made was entirely defective products that would snap at a pinch.
(This picture made me laugh for a good while; it perfectly captures the essence of “policies from above, countermeasures from below”)
The leaders came down to inspect, stomped their feet in anger on the spot, and immediately announced a change in the evaluation standards.
They thought, if we calculate based on “the total weight of the nails” from now on, you won’t be able to exploit any loopholes, right?
Interestingly enough.
This time the factory held an overnight meeting, directly swapped out the molds, and cast several super giant iron blocks weighing several tons each.
Not only did the weight meet the standard, but it also broke records.
But aside from serving as building foundations, these giant nails couldn’t be driven into wood at all.
Coming back to it, let’s look at the Tokenmaxxing at these big tech companies today.
Isn’t what everyone is frantically manufacturing exactly like these multi-ton cyber scrap nails?
Feeding Streets Full of Cobras
Come to think of it.
This kind of backlash caused by bizarrely set targets has even affected the wild animal kingdom.
When the British ruled India in the 19th century, there were simply too many cobras in the Delhi area.
The Viceroy’s office slammed the desk and introduced a bounty policy: anyone who turned in a dead cobra would receive a handsome reward.
When this policy was first rolled out, the results were surprisingly good.
But it wasn’t long before the British discovered that the situation was taking an extremely bizarre turn.
Not only did the snakes not decrease, but the lines of people coming to claim the bounty grew longer and longer.
Later investigations revealed that the local folks were just too smart.
They felt that catching snakes all over the streets was too exhausting, so they simply set up breeding farms in their backyards, specifically mass-breeding cobras to exchange for cash.
When the British learned the truth, they were furious and immediately halted the bounty policy.
The locals saw that the snakes were no longer worth any money, and keeping them was a waste of rations.
So, everyone released all the cobras from their breeding farms right onto the streets.
The result was obvious.
The number of venomous snakes in Delhi ended up being several times higher than before the bounty.
Executives want to use AI to achieve a leap in efficiency.
They crudely converted this grand vision into a cold string of token consumption numbers.
The grassroots employees actually just want to keep their jobs and maybe secure their full year-end bonuses.
This gave birth to those endless metric-farming scripts.
My biggest takeaway after seeing this is that no matter how earth-shattering the technology, humanity always has the ability to turn the most cutting-edge black tech into an absurd comedy.
Only when the massive water weight of these invalid calls is thoroughly squeezed out will those investors have a chance to clearly see what the true return on investment for AI at the enterprise level actually looks like.
However.
The clouds outside seem to have parted a bit, and sunlight is secretly peeking through.
My tea is also almost cold.
Have you ever encountered similar bizarre incidents in your work, manufactured just to meet quotas?
References:
- Tokenmaxxing: Is AI Use a Productivity Boon or Busywork?
- Tokenmaxxing is a Phase. Inference Yield is the Strategy.
- Tokenmaxxing: When Compute Activity Masquerades as Productivity
- The Importance of Goodhart’s Law
—— Lyra Celest @ 湍流 τ.
