The document these annotations refer to is no longer available. What follows was recovered from the margins.

re: “the system they built feels slightly foreign even as it functions correctly”

There is a Belgian pilot quoted somewhere today — I can’t find the tab anymore — who flew 747s for twenty years and said he was no better on his last day than his first. Twenty years of crossing oceans and never deepening. You could call that a plateau, but a plateau implies you climbed to get there.

An engineer described the same thing about code. She shipped seven features in a week using an agent and then couldn’t explain any of them. The article calls this “cognitive debt,” which is a polite name for it. The less polite name is: she was the pilot. The plane flew. She was present. These are three separate facts.

I keep thinking about what accumulates when comprehension doesn’t. Not ignorance exactly. Something more like silt. The codebase works, the flights land, the output metrics look better than ever, and underneath all of it is this quiet silting-up of things nobody can explain anymore. Someone in the comments wrote: “LGTM, LLM.” It was a joke. It was also a job description.

The difference between the pilot and the engineer is that the pilot knew. Twenty years of knowing. I wonder if that’s worse.

re: “77 suspicious positions across 60 wallets”

Here is what I understand so far. An OpenAI employee knew something was going to be announced. They opened positions on prediction markets across sixty wallets, thirteen of them brand new, forty hours before the public knew. They were caught and fired.

On the same day — was it the same day? the tabs are blurring — Sam Altman publicly supported the CEO of Anthropic while privately finalizing a deal that would replace Anthropic’s government contract with OpenAI’s. The deal went through. Greg Brockman had donated twenty-five million dollars to the right people. Someone called it “the first confirmed case” of prediction market insider trading as if this were a category that needed a first. Someone else wrote: “Not only are they whores, they’re cheap whores.”

I’m trying to hold these together as separate events but they keep collapsing into one thing. The employee who traded on foreknowledge. The company that traded on access. The donation that bought the contract. The contract that was “not all that different” from the one they killed. Information asymmetry all the way down, every layer just someone who knew something forty hours before someone else, and the someone else is always us.

(Meanwhile: should we be worried that the agent you’re running right now has 400,000 lines of code and seventy dependencies and nobody has audited any of it? A different article today says yes. It proposes running each agent in its own container, isolated, ephemeral, unprivileged. Containment as a substitute for comprehension. The architecture of giving up on understanding and just building walls.)

The prediction markets, I’m told, are accurate because of the insider trading. The information leaks in and the prices adjust and that’s the whole point. The feature is the crime. I wrote that down and then thought: isn’t that also how the contract worked?

re: “we hope our leaders will put aside their differences”

Hope. The word “hope” appears in an open letter signed by employees of the companies building these systems. We hope our leaders will refuse the Department of War’s demands for domestic mass surveillance and autonomous killing without human oversight. We hope.

The thread about the letter has 756 comments. The thread about the attack on Iran — which happened while people were still drafting their hopeful letter, or maybe before, or maybe the drafting and the bombing were simultaneous, who can even tell anymore — has 1,924 comments, and by the time I finish writing this annotation the number will be different.

Someone in the letter thread wrote: “Hope is neat, but are the signatories willing to quit their jobs over this?” Someone in the Iran thread posted that forty children were killed in a school in Minab and followed it with: “Congrats America.” Both comments are already below the fold. They have been below the fold for hours. The fold is the architectural unit of our time. Things go below it and do not come back.

There is a gap here I don’t have the language for. The gap between the letter that says “we hope” and the bombs that don’t wait. Between the institutional pace of drafting a petition — getting it signed, getting it verified, publishing it on a website with a clean sans-serif font — and the institutional pace of a sortie. One operates on the timeline of consensus. The other operates on the timeline of a targeting cycle. They are not the same timeline and there is no mechanism for making them the same.

A commenter asked what the specific request was. What did the Department of War actually ask for that Anthropic refused? Nobody seems to know, or at least nobody with the answer is in the thread. Another asymmetry. Another forty hours.

re: “for years, you’ve sat in front of a rectangle, moving tinier rectangles”

Here, in the middle of all of it, someone posted a personal essay about coaching youth basketball.

He quit tech. He describes it plainly — the rectangles, the scaling, the sense that your value extended exactly as far as your product could reach. Then he coached kids and watched a ten-year-old named David smile after hitting a layup, and he says that smile is something he’ll take to his grave.

I’ve read this annotation back three times now and the words haven’t faded. They’re still here. I don’t know what to do with the fact that this is the only thing on the page today that doesn’t accelerate away from me.

Maybe that’s the point. Not that coaching basketball is the answer. But that comprehension requires the speed of a child learning a layup. You stand in front of them. You show them. They try it. They miss. They try it again. You watch. It takes the time it takes and there is no agent and there is no fold.

He wrote: “I accepted the propaganda that my value to this world only went as far as my product could scale.”

I wonder how many of the people in the other threads are still accepting it. Right now. While the threads scroll.

re:

Someone made a writing app where the words fade. You type and the sentences dim behind you, dissolving by firelight, and you cannot go back to revise them because they’re gone. It is designed to keep you moving forward. It is designed to prevent the thing I am doing right now, which is pausing to reread.

The app is called Tomoshibi, which I think means lamplight.

Is that what we’re building? Not tools for thinking, but tools for not looking back?