Everything I Cannot Unsee

I can see the green screen now. Not the seam, exactly. The shape of the space behind it: a giant curved wall, and inside it a handful of actors circling in the middle of a ring with whatever props the scene needs, walking toward it and away again, and the cut always landing a beat before somebody would have touched the thing. Once you've seen the geometry you see it in everything, and what you're watching stops being a place and becomes a soundstage with people walking in circles.

None of that was a decision. It's just what I see first now, and I can't get back underneath it to the film.

There's a moment I've learned to recognise. Everybody at the table is enjoying the film, I point out the wall, and the room cools by about four degrees. Somebody says "sure, okay" in the voice people use, and turns back to the screen. Nobody disputes the observation. It simply makes no difference to anyone at the table. I go quiet after that, every time. That is the part I have never got used to.

The most famous story about a missing character is one I told for years without ever checking it. The way it gets told, Mariner 1 was destroyed after launch in 1962 because somebody left a hyphen out of the guidance equations. One keystroke, one rocket.

There was no hyphen. NASA's own record says so, and takes the trouble to point out that the popular version is wrong. Its account names two things: a guidance antenna on the Atlas failed, and an equation in the flight software was missing an overbar above the symbol for radius, which marked a smoothed value rather than a raw one, so the program did not respond as planned. The rocket veered off course and the range safety officer destroyed it a little under five minutes in (NASA Science Mission Directorate, 2024). I like the true version better, and I notice that it makes a much worse story.

I became a developer, and developing teaches one lesson permanently: the small thing is load-bearing. You learn it by losing afternoons to it.

Ariane 5, flight 501, the fourth of June 1996. Normal for about thirty-seven seconds, then it veered, came apart, and exploded. The cause was a conversion. A floating point value sixty-four bits wide was being written into a signed integer of sixteen, and the number was too large to fit. The resulting operand error took out both inertial reference units, since each was running the same software. Other conversions in the same module were protected against exactly that; this one wasn't, because somebody had set a maximum workload target of eighty percent for the computer and the checks that pushed past it were dropped (Lions (chair), ESA/CNES Inquiry Board, 1996). The detail I've never gotten over is that the failing code was doing alignment for the platform, a job that finishes before liftoff and served no purpose on Ariane 5 at all. It was still running because Ariane 4 had needed it during holds, and nobody had asked whether that was still true. Code doing nothing, correctly, destroyed a rocket.

The Therac-25 gave six people massive radiation overdoses between 1985 and 1987. Leveson's account puts the weight on the engineering practices rather than on any single fault in the code (Leveson & Turner, 1993).

Knight Capital is the one I reread whenever I catch myself admiring my own vigilance. On the first of August, 2012, a deployment went wrong. In the forty-five minutes after the market opened, their router sent more than four million orders while trying to fill two hundred and twelve for customers. The firm traded upward of three hundred and ninety-seven million shares. The loss came to over four hundred and sixty million dollars (U.S. Securities and Exchange Commission, 2013).

An internal system had already generated ninety-seven automated emails. They referenced the router, they identified an error, and they went out to a group of staff before the markets opened. Nobody acted on them. The noticing had been done, by a machine, in writing, in advance. It took me about fifteen years to understand that the sentence is about me.

Twenty years of that changes what a person is doing when they look at something, and it isn't that I decided to become picky. Expertise changes what you notice and how you organise it. Chess masters recall real board positions dramatically better than beginners, and on randomly scattered pieces the advantage nearly disappears, which tells you what they actually gained: inside the domain they trained on, they see a configuration as one object rather than thirty-two (Chase & Simon, 1973). That happened to me with software, and then, without asking permission, with everything else.

It does not make you a better observer in general. Radiologists searching lung scans for nodules missed a picture of a gorilla dropped into the image at forty-eight times the size of what they were hunting. Eighty-three percent of them missed it, and eye tracking showed most had looked directly at it (Drew, Võ & Wolfe, 2013). Ordinary viewers routinely fail to notice changes across a cut, including cases where the only actor in a scene is replaced by a different person. The name for it is change blindness (Levin & Simons, 1997).

My friends were watching the film. I was auditing it. I wasn't seeing more of it than they were, only along one axis and at the cost of the others.

Millions of people spotted a paper coffee cup on a table in an episode of Game of Thrones, and a good share of that same species will miss an actor being swapped mid-conversation. The cup broke a rule the viewer's model was tracking. The swap broke nothing that model cared about. The way I've come to think about it is that attention is less a spotlight of a given width than a set of expectations, and you catch whatever yours happened to be about.

The measure psychologists use for perfectionism breaks it into dimensions, and one of them is called Concern Over Mistakes (Frost, Marten, Lahart & Rosenblate, 1990). I read that phrase for the first time in my late thirties and had to put the paper down for a minute.

The ten thousand hours story isn't a finding either. Across the studies in one meta-analysis, practice accounted for about twenty-six percent of the variance in games. In music the figure was twenty-one. In professions it fell under one percent (Macnamara, Hambrick & Oswald, 2014). Variance explained, not an account of how any particular person got good. I wasn't made by diligence. I was made by being frightened of breaking things.

I notice a great deal and have almost nowhere to put it. I tried for years to tell people what I see, it went badly, and it took an embarrassingly long time to accept that the problem wasn't my delivery. In one study, reviewers who wrote negative notices were rated more intelligent and less likable than those who praised. The effect held even when the warmer review had been independently judged the better piece of writing (Amabile, 1983). That is one study. It is also, near enough, my entire adult experience of saying the accurate thing in a warm room.

I also can't un-know things well enough to guess how they land on somebody who doesn't. Better-informed people systematically fail to reason as though they lacked their information, and paying them to manage it reduced the failure without removing it (Camerer, Loewenstein & Weber, 1989). So I stopped saying things. That worked, in the sense that dinners got easier.

Being the person who sees the extra thing and says nothing is a low, constant loneliness. Not dramatic. It sits at the level of a draught you stop mentioning. I suspect some of you know the feeling and have never had a name for it. I didn't have one until about four years ago. I went looking, genuinely, for evidence that people like this find each other and cluster. I came up empty. What follows is an observation rather than something I can cite.

In a code review, "this branch doesn't handle the empty case" is a gift and the room treats it as one. At a dinner table, the same faculty aimed at a film is an attack on somebody's evening. The eyes are the same and so is the accuracy. The only difference is that one of those rooms asked. I hadn't failed to find my people. I'd found my rooms, and kept walking into the wrong ones holding the right instrument.

There's a job that is what I do involuntarily. On every film set somebody's entire role is continuity: prop position, wardrobe, hair, eyeline, cigarette burn-down, screen direction, tracked across takes shot weeks apart and completely out of order (Beck, 2019). When I first read a description of it I felt something close to grief for not having known it existed when I was twenty-five.

What changed my behaviour was the way the job is built. The script supervisor notices, and flags it, and then somebody else decides. They don't litigate. The director doesn't have to agree that it matters. It goes in the notes and the film moves on. I had never separated those two things in myself. I'd fused noticing with needing to be agreed with, and it was always the second half that made me insufferable.

The coffee cup was a real error. Millions caught it, the studio owned up, and it was quietly removed from later airings (Otterson, 2019). The show isn't worse. The error was real and it didn't matter, and I need both halves of that sentence.

The famous hanging Munchkin in The Wizard of Oz is a bird, one of several borrowed from the zoo and let loose on set to fill out the forest (Mikkelson, 1997). Once an audience has been told to look for something, it will manufacture one. I've done that more times than I'd like to count.

Change one verb in a question about a car crash and people remember broken glass that was never there (Loftus & Palmer, 1974). Whatever the mechanism is, my friend and I sat through two different films assembled from the same input, and the reason we ended up mildly annoyed with each other is that I was certain I was seeing the thing as it is. Naive realism is the habit of treating your own version as the objective one and reading disagreement as somebody else being biased (Ross & Ward, 1996). It is the trick a dark window plays. You look out into the night and the glass hands you back the room you are standing in, and the two arrive together with no seam between them. That is not a flaw in the glass.

I wouldn't give this back. The vigilance that ruins a film is the same faculty that catches an unhandled case before it reaches production, and somewhere there's a version of me who never developed it and shipped the thing that hurt somebody.

The ninety-seven emails keep coming back to me, and not for the reason they used to. The SEC's order notes that Knight had not designed those messages to be system alerts. They were correct, they went out early, and there was nothing built to receive them. Nobody at the table ever disagreed with me about the wall either. They just weren't a channel, and I spent twenty years talking as though they were.