Cosplay Epistemology
Going through the motions is not the same as getting in the reps.
I am currently sitting in a cafe-bar in Korčula, an island off the Dalmatian coast of Croatia, retracing the route Rebecca West took in Black Lamb and Grey Falcon1 — and let me tell you, this region has got it goin’ on. The perfectly-crisped fish on my plate was swimming an hour ago. The streets are paved in stone worn smooth by eight centuries of feet. My wife has threatened to charge me a gelato for each new adjective I use to describe how blue the water is.
Yesterday I took a small boat across the channel to Badija, an island that contains nothing — no village, no shop, no road — except a fifteenth-century Franciscan monastery and a colony of unbothered deer2. I sat in the cloister for an hour and listened to the sound of water pouring itself into crashes of rock.
Not bad, in other words.
But still, it’s Europe — meaning that what hits me even louder than any of that beauty is all the damn cookie popup banners in my browser.
Every website. Every single one. Would you like to accept cookies? Manage your preferences. Accept all / Reject all / Some unmarked third option that takes you to a six-tab settings panel governed by a Latin acronym you have never heard of. I am trying to look up the ferry schedule. I am trying to find out whether the konoba down the road is open on Tuesdays. I do not, in the moment, have rich opinions about how Vimeo handles my behavioral analytics. I would just like to use the website, please, thank you.
Making things worse is the reality that if you click “reject,” half the time the site stops working in some mysterious and difficult-to-undo fashion — which means the choice on offer is not a choice, but a clickthrough disguised as one.
Sure, there is a case to be made about not using cookies. But in practice I’d obviously like to use cookies (and I’d definitely like to avoid a clickthrough). Because what matters most is the part where the product works the way it’s supposed to. Not the part where I fixate on any particular set of hoops to jump through first.
Increasingly, the choice about whether or not to adopt AI is playing out like that.
Even when it comes to pizza.
A couple of days ago, Axios reported that Andy Brown — DJ turned restaurateur, world-champion pizzaiolo, founder of Andy's Pizza — has spent the last year building a custom operating system for his fourteen DMV locations and roughly 350 employees.
Axios describes it as “AI-powered.” This is true in a narrow sense, but sort of misses the zinger.
The AI does not run the pizza shop.
The AI let Andy rebuild the pizza shop.
As orders roll in at 6pm on a Friday, the math of whether you are about to get buried under an avalanche of demand is, in theory, straightforward. A perfect stone-oven pizza takes six minutes, a reheated slice takes four, and Andy has 24 stone-oven slots to work with. Take too many orders during the wrong fifteen minutes and you spend the rest of the night playing catchup.
The thing is, “straightforward” math still requires someone to do it flawlessly. That’s easier said than done when you’ve got a shop full of hungry customers.
Now — courtesy of an apparatus Andy built himself with Claude Code — the software does the math. The shop automatically turns orders off before it gets buried. “A huge sigh of relief to our team,” Andy wrote. “When you guess wrong and take too many orders — you fall behind which means you’re behind all night. Super stressful.”
Andy is not a coder. He has, by his own account, broken things and rebuilt them from the ground up. But the whole thing just flows better now that they’ve worked out the kinks.
And the vibe’s still good. “None of what we done replaces people,” Andy takes care to emphasize. No kiosks. No QR codes. No creepy-AI-voice phone orders. My go-to slice of margherita still costs $4.50. But, Andy says, “it takes SO MUCH mental load off when you don’t have to look up at a bunch of tickets.”
This is what AI deployment actually looks like when it’s working. Andy used the robot to write some Python that solves a known problem that no existing piece of software could address. He used it where it earned its place.
The same shape also comes up in the social/civic sector — take tools like GroundVue, or Abigail Haddad’s Federal Contract Terminations Tracker. An operator of some sort has a long-standing problem, and is looking at a bunch of data that’s relevant to the problem but which has been prohibitive to process for whatever reason. The frontier model’s job is to make operationalizing that information possible, often for the first time.
With something like GroundVue, such information comes from public meetings. In Andy’s case, the info comes from the kitchen. There are a million ‘ground-level’ use cases that work this way.
And then, by contrast, there’s “tokenmaxxing”.
The Financial Times reported earlier this week that Amazon employees have started using an internal agent platform called MeshClaw to automate non-essential tasks — not because anything actually needs automating, but because the company set a target of 80%+ of developers using AI weekly, and is tracking token consumption on internal leaderboards. Hence the ‘maxxing’ of the ‘tokens’.
As Jordan Weissmann observed, this is ironic coming from a company known for its disdain for vanity metrics.
Doug Colkitt put it a little more vividly.
Amazon is hardly alone in this. A Meta employee built an internal leaderboard called Claudeonomics before Meta killed it when the story leaked. Nvidia’s Jensen Huang has publicly said he’d be “deeply alarmed” if a $500K engineer wasn’t burning $250K in tokens3.
The exquisite Han Lee, whom I’ve been following since his World of Warcraft days and who now leads the machine learning effort at Moody’s, has dubbed this phenomenon the AI Great Leap Forward.
In 1958, Mao ordered every village in China to produce steel. Farmers melted their cooking pots in backyard furnaces and reported spectacular numbers up the chain. The metal was useless — brittle pig iron no factory could work with — but the reports flowed upward and the cadres who filed them were promoted.
Meanwhile the grain in the fields rotted because the people who should have been harvesting it were tending furnaces. By the time the lie ran its course, thirty million people had starved.
The structural error is identical to that of the dashboards: solving for what looks like the thing instead of solving for the thing. Mao’s villages were performing industrialization. The smoke was real. The furnaces were real. The reports up the chain were real.
It’s just that the steel was not.
The token is not the work. The token is the smoke from the furnace.
Tech rollouts are hard. Deploying something as comparatively straightforward as Slack or Salesforce or Notion is often a multi-month nightmare, to say nothing of a vanguard technology that changes the nature of work itself. Trying to speedrun all that with a mandate or a leaderboard is just pushing on one end of a noodle.
Ethan Mollick made this point earlier this week from a different angle.
As long as people are required to figure out how the technology fits inside an actual organization, there’s a job to do. The Accentures of the world are still going to be in business. The hard part is the part where you figure out how the tool gets you the result you need.
Andy Brown, sitting in DC on an off-the-shelf Max plan, is the forward-deployed engineer for Andy’s Pizza. He is the integration consultant. He is the systems architect. There is no Accenture in his way, because the cost of the work has collapsed to the point where the operator who knows the problem (i.e. him) can do the integration himself.
Andy’s approach passes muster because it isn’t coming from a place of fear.
He isn't afraid of being left behind. He isn't afraid of being replaced. He isn't afraid of being the guy who didn't see what was coming. He is a guy who runs fourteen pizza shops and got tired of jamming his business into software designed for online t-shirt vendors. So he sat down with Claude Code on the Max plan, broke a bunch of stuff, rebuilt it from the ground up, and can now get out ahead of the Friday crunch. He knew what the work was, because he had a job to do.
Tokenmaxxing is the opposite. It's what you do when you don't have a problem to solve but you've been told you'd better look busy anyway.
It's the same thing we see in the social/civic sector all the time: outcomes are not the same as activities. The actual outcome you want is murky and multi-factor and complicated to parse, so the temptation is to create vanity metrics around what you can measure and can prove. It’s a form of ‘legibility cope’: artificial rigor papering over an intrinsically difficult-to-model question.
This whole phenomenon embodies what I like to call Cosplay Epistemology: the belief that the performance of a thing is a version of the thing itself. And it fails for the reason the prototypal cargo-cult example fails: a conflation of cause and effect.
In Cosplay Epistemology, the mask fits. The vocabulary is right. The token counts get tracked and visualized. The badges say Token Legend and Cache Wizard and Session Immortal, and the “AI strategy” gets socialized successfully at the offsite. None of it touches the mechanics of any actual problem, because the problem has been defined as the performance. The pop-up banner alerts you to the hoops you need to jump through, and says nothing about whether or not the product works.
Anyone doing important work has a responsibility to do it as well as possible. That means rejecting the urge to go full Luddite. But it ALSO means rejecting the urge to bush-league ‘e/acc’ via clown-shoes performativity and jumbotron stat-maxxing and every other variant of backyard fake-news ‘steel production’.
Andy didn’t deploy AI to replace the line cook. He used AI to build the system that lets the line cook stop staring at a wall of tickets and start paying attention to the pizza.
The pizza is still made for people.
The apparatus that helps make the pizza is the new thing.
That’s the part worth doing.
The rest is smoke.
For the unfamiliar: scoped in 1937, finished in 1941, ~1,200 pages, arguably the greatest travel book of the twentieth century and definitely the most ambitious. Imagine a vintage-hot Fabian aunt with the discernment of Henry James, the heart rate of a war correspondent, and less interest in being concise than my Starcitygames.com Magic column. I am, predictably, obsessed. The epigraph? “To my friends in Yugoslavia, who are all now dead or enslaved.”
Unbothered, that is, until I got there <puts on sunglasses>
The vast majority of those tokens, not incidentally, get processed on Nvidia chips, which I might mention (given my Rebecca West kick) ‘does not smell entirely of the rose’.