Optimizing the Point Away
Nothing in Minecraft tells you to punch the tree.
Nothing warns you that night is coming, or what the crafting table does, or that the green thing walking toward you is about to explode. The game was, for practical purposes, unplayable without another kid who already knew. Everything you needed lived outside the game, in the schoolyard, the computer lab, the wiki, the older brother, passed hand to hand like it was contraband. And it became the best-selling game in the history of the medium.
Measured against the last forty years of playtests, funnels, retention curves, and postmortems, Minecraft is a catalog of errors: a brutal, unexplained first night with no tutorial, no goals, and no help coming. The most successful product the medium ever produced fails the entire accumulated curriculum, and it isn't a miracle, so the curriculum is wrong. The tempting save is that Minecraft's core loop was simply so good it could pay friction's tax. But the friction wasn't a tax on the loop; it was in the loop. The withheld knowledge is what made another kid necessary, and the other kid is what made it spread.
I want to argue something stronger than the curriculum being wrong, though: the industry, tech and games both, has been forgetting how to build great things while believing it's been learning to build them better, and the learning machinery is the mechanism behind the forgetting.
Forgetting at n=1: what happened to Viva Piñata
You can watch it happen inside a single studio, four years, one franchise. I recently replayed the original Viva Piñata and its sequel on the same day.
The first thing you do in Viva Piñata is walk down a beautiful hill into a ruined garden and talk to a crying girl standing in the rubble. She hands you a shovel and asks for help cleaning up the mess. You start by beating cracked ground into dirt. Effort, violence against the decay. Then a small worm appears, colorless, because the dirt you just made is its habitat. It moves in as you keep working, and then it blooms into gorgeous technicolor. In twenty minutes the game has taught its entire thesis without a single rail or forced action: life in this world is downstream of your work.
The sequel throws you into a pretty garden decorated with grass, flowers, and stepping stones. A creature lies penned in a fence. The first new one appears before you've done anything (yes, in an unskippable cutscene before your first action; I wish this point were harder to make), and within seconds the game assigns you a forced task. Life in this world is content to click through.
Nobody at the studio experienced these changes as mistakes. They were "improvements", by every measurement the studio could make. The empty start almost certainly playtested as friction: players took too long to reach their first piñata. So the garden got pre-filled. Nobody in the room knew what the empty garden was for, nobody had the will to defend it, and no one came to save the whole point of the game from optimization.
Metrics kill games by making fun inadmissible
The industry has what it sees as a learning system: playtests, A/B tests, dashboards, postmortems, conference talks, retention funnels, churn. The system can only learn what it can measure, so anything it can't measure stops counting as an argument.
To be clear, most friction really is just bad UX, and it should die. The disease is that the instruments can't tell the difference: confusion that is the game and confusion that's a broken door report identically, so both get removed.
Friction's costs are visible to everyone at the table:
- churn
- funnel drop-off
- retention as a percentage
Friction's benefits are invisible until they're everything:
- ownership
- folklore
- a sense of wonder
- the story someone tells in ten years
- the friend who teaches you the secret
These are all long-term properties. None of them can be seen inside a playtest or a quarter, and up front they cost you slower adoption. So every optimization-driven decision removes them, and the door only swings one way, because few producers ever green-light restoring confusion (the good ones always do). Each removal the data forces is locally correct, and the sum is a game nobody loves.
The filter also runs on the people. Defending illegible value is a career risk. A producer who restores confusion has nothing to show for it at review season, so a professional can't keep making bets their own instruments call insane. The only people who can are the ones building because they love the thing. Carmack and Notch never asked a dashboard for permission; the conviction was the company. Everyone else forms their taste inside the previous generation's already-smoothed games, until the tooltip layer stops being a choice and becomes what games are. Between broken-up teams, only the articulable transmits: systems, content cadence, retention tricks. The tacit part dies in every handoff, and four generations later you're shipping photocopies of photocopies.
The proof is what happens when someone defects and wins. FromSoftware spent fifteen years ignoring onboarding wisdom, from Demon's Souls through Dark Souls to Elden Ring: obscure, unexplained worlds that force communal sensemaking through wikis, ghost messages, and the friend who knows the secret. The bet kept paying all the way up to Elden Ring sweeping game of the year and outselling nearly everything, with no quest markers and barely a journal. The industry still couldn't absorb the finding. Its learning system extracted "difficulty," the one measurable ingredient, and shipped a decade of hard games with quest markers. Plenty of individual designers know the old wisdom: Outer Wilds exists, and about twenty people made it. The industry has no organ that can hold it.
You can't trust the metrics you think you can trust
Take the most honest number anyone has: latency. It isn't a proxy, it's barely gameable, and users genuinely want it low. Nobody's dark pattern ever made a game faster.
Chasing it is how MMO worlds went static. NPC schedules, physics, and persistent simulation don't fit the tick budget, and "computationally expensive" is nearly a synonym for "interesting." The worlds got faster and stopped being alive. The web hit its Core Web Vitals the same way, by becoming skeleton screens and deferred husks: the page paints in 800 milliseconds and contains nothing.
The defect has nothing to do with proxies. A metric specifies what to achieve while staying silent about what may be sacrificed to achieve it, so even well-intended optimization ends up spending whatever the metric doesn't price.
There's a timing corollary I hold for products generally: metrics are trailing indicators of feeling at scale. Once something is loved, a dashboard has real uses. It makes the feeling legible to investors, and it can stop a thousand new hires from wrecking the thing quickly. But that's the whole job, a guardrail against fast destruction, and if you stop measuring the feeling yourself, the guardrail starts steering. Pointed at a product before the feeling exists, instruments measure the legible fraction of noise and drag the product toward whatever that fraction rewards. Measurement can protect a known good, but it can't discover one. Discovery happens somewhere else, in taste, conviction, and watching five strangers' faces. The industry's disease is running the protection toolkit during the discovery phase, permanently, at every studio at once.
I run a game platform, so none of this is spectator opinion. These are rules we live under, and they cost us regularly. Spawn ships mysteries we refuse to document, and our players build the folklore, including a confidently wrong method that circulated for a day before the community tested it. In playtests the house rule is to stay quiet and let people struggle, because directing them skips the pain, and the exercise exists to find the pain. We don't have a dashboard. Maybe this costs us the company. It would still be the right price; a good bet that doesn't pay was still a good bet.
The machine that studies harder
And more diligence makes it worse. Every additional playtest, every finer-grained funnel, every well-run postmortem pushes more decisions through the filter that strips the important part.
Don't expect a fix. I've lost this argument at work more times than I can count: I say the confusion is the game, someone cites the playtest, and a manager who doesn't care sides with the number, because a number survives being forwarded to someone who doesn't care and taste doesn't. Even physics couldn't beat this. Mercury's orbit falsified Newton for decades and the field filed it as noise, invented a planet named Vulcan to explain the drift, and went looking for it; the correction came from a patent clerk the hiring filter had already rejected. In every case where the point survived, someone had the power to skip the meeting. Notch owned the company. Miyazaki is the president of his. Bridges fall down, so bridge-building self-corrects. A game that lost its soul doesn't fall down. It sells for a decade on momentum while the point leaks out through churn nobody can attribute.
Games are only the leading edge. They're the most feeling-dense product category on earth and also the most instrumented, so the rot is most complete there first, but you can see it everywhere, including in the aesthetics of the world, which are dying of a similar disease. The same machine runs everything. Including the parts that were meant to make you feel something, and now can't.
In 2010, Minecraft generated a small wooden house for you at spawn. Players called it the Indev House. It was the optimizer's answer before the industry could deliver it: shelter provided, first night solved, friction removed.
Notch deleted it and saved the whole point of the game.