What are you if not your memories?
The way your mother’s voice sounded calling you in for dinner. The smell of a place you haven’t been in twenty years. The particular weight of someone’s hand in yours the first time you held it there. The moments with friends you’ll replay until they become folklore. How proud you felt the day you finally pulled off something you had been working toward for so long, exactly where you were when it happened, what song was playing, what the light looked like. And the strangest, most miraculous thing is how thin the membrane is between then and now; how a snatch of a song on a stranger’s speaker, a smell drifting out of a doorway, the sound of someone’s laugh, can pull you under so completely that for a moment you are not remembering, you are there, living it all again.
And long after you’re gone, your loved ones’ memories of you will still live in the space between them, and they too will be pulled under one day, and they’ll reach out in that void and find you.
When I wrote the words above, I sat on a bench in Nakameguro, underneath the sakura, when the petals were coming down so steadily you could watch them collect on the surface of the canal. It was the kind of afternoon you don’t quite earn; you just happen to be present for. A woman near me was reading a book she’d clearly read before. An older man stopped to photograph a tree he must have photographed a hundred times. There is something about that place in that week of the year that makes everyone slow down at the same time, as if the city had agreed, briefly, to remember itself.
I put the pen down and took it in. I have built startups and software for most of my life, and for a while now I had been somewhere else, doing things that did not pay and did not scale. What I was thinking about, on that bench, was the difference between the old world I had built things in for most of my career and the new world I had been living in since. The old world had not been perfect, but it had been recognizable. People still had to learn things the slow way. The work of doing something difficult still belonged, mostly, to the person doing it. The new world was already something else.
A kitchen in Thailand, where the line cooks threatened to stab me in Thai when I couldn’t lay an egg net over the pad thai on my first try, and the woman who ran the place fed me before service every night without asking what I wanted. A photo booth I rebuilt with chemicals and timing belts and put in my living room, where friends and strangers crammed in and pulled the curtain and came out with a strip of pictures of themselves. Films I helped make. A body of work I shot and hung on a wall. Dozens of other rooms I was lucky enough to stand inside of.
None of these endeavors had anything to do with each other. A film set and a wok station and a darkroom and a living room with a photo booth in it are not the same room. The crews don’t overlap; the vocabulary doesn’t transfer. And yet, across all of them, the same thing kept happening.
Someone would hit a small wall: a translation, a research question, a creative block, a logistical snag, a name they couldn’t remember. And someone else in the room would say, just put it in ChatGPT. The answer would come back confidently wrong, half the time nobody would check, and the project would carry the wrong answer downstream until it broke something.
Over the years, what had been an occasional shortcut became, for a lot of people, the default.
And so it was in that moment on the bench that I understood that nobody else was going to fix it. A great many people were going to claim they could be fixing it, and almost none of them were going to mean it, because the business model of fixing it doesn’t look like the business model of the thing that’s broken.
A week later, I founded 6OVER3 INSTITUTE.
The things I’ve seen, and that we all will continue to see, are pointing to a problem with the shape of society’s soul.
Thinking is being outsourced. Every generation hands something off to the next tool — the calculator, the spell-check, the map app on the dashboard — and every generation survives it. This time it’s different. We are trading the act of learning, of struggling with a problem long enough that something in us reshapes to fit it, for immediacy. To be placated. The difficulty of doing the thing was how the person was made.
Generative AI, in any of its flavors, whether code, text, image, or video, is a lossy compression of every original thing humanity has ever made, replayed at you in real time. It produces work that looks, feels, and, crucially for software, performs roughly the same as everything else it was trained on. So you babysit it. You correct it because it cannot see what you see. You burn millions of tokens and hours, sometimes days, trying to coax it toward a vision it has no access to. I keep thinking about self-driving cars. Can it sometimes get from A to B? Sure. Is it cheaper, faster; does it know the shortcuts you know? Not really. The car is a tool. Driving it, applying what you know to the road in front of you, is the skill. We are at risk, on a civilizational timescale, of forgetting the difference.
The frontier labs build the models, the new picks and shovels. Everyone else builds the products that route your thinking through them. The aggregate effect, given enough time, is that everything in the world begins to look and sound and feel the same. The texture flattens. The accents go. The strange little decisions that made a thing feel like it came from a particular person in a particular place at a particular time get smoothed out into a median that pleases no one and threatens nothing.
Training data is the oil of this industry, and the easy wells are dry. OpenAI was quietly pumping the surface deposit, the open web, for years before anyone outside the field had heard the company’s name. Once OpenAI made it obvious how much there was to extract, everyone else piled in. Nobody had yet decided that asking was the convention. That surface is mostly gone now. The open web that produced it is collapsing; the thousands of third places that used to hold the actually valuable information, the forums, the message boards, the personal sites run by people who knew one thing well and wrote about it for the love of it, have funneled into a handful of platforms that now claim ownership over what their users post and license it back to the model companies by the petabyte.
What’s left on the surface, increasingly, is the runoff of the machines themselves: a slick of synthetic text and synthetic images that the frontier labs now know they cannot reliably filter out of their own scrape. A trained human reader can usually tell. Their crawlers can’t, outside of a few telltale watermarks that almost nobody bothers to embed. So every new pass over the web pulls in more of the previous generation’s output; the next model trains on its predecessor’s exhaust; the feedback loop tightens. The well is recycling its own waste, and the people running the pump are the ones who poured the waste in.
So they’ve moved to fracking.
The next reservoir, the one being drilled into right now while you read this sentence, is you. Your private notes. Your half-finished drafts. Your designs, your voice memos, the journal entries you’ve kept in some cloud service for ten years because where else would they live. Pressurized, hard to reach, and, until recently, considered off-limits by anyone with a working sense of decency. They’ve decided it isn’t off-limits anymore. They’ve decided it’s the resource.
If you’ve used a service for years to store your private work, there is a real chance they flipped a switch in the last year and began training on it. You have to opt out, because they know that nobody, when you put it to them in those terms, would willingly opt in. It used to be common sense that if the service was free, you were the product. Now you pay two hundred dollars a month, and you are still the product.
Even if you opt out, “private” is doing less work than it sounds like it is. What you put into someone else’s servers passes through Trust & Safety teams, engineers, third-party processors, government agencies, and, increasingly, other models, quietly classifying, tagging, routing, flagging it for review. Once it’s uploaded, it isn’t just yours anymore.
And like fracking, the damage isn’t only to the thing being extracted; it’s to everything around it. The water table. The bedrock. The people who happened to live above the deposit. When a company decides your decade of private writing is a resource, it isn’t only your writing they’re changing the value of. They’re changing what it means, for everyone, to write something down and not have it read.
I believe AI can change the world for the better. But I don’t believe the current trajectory is going to get us there. By virtue of the way it’s being built, who owns it, what it’s trained on, where the compute lives, what it’s optimized to do, the path the industry is on cannot produce the future it keeps promising. I don’t think the current generation of LLMs, or the line of work that produced them, is getting us any closer to AGI. We are scaling a kind of intelligence that gets more fluent without getting more present.
The dominant vision, the one being sold to you in keynote after keynote, is centralization. A handful of enormous models, in a handful of enormous data centers, owned by a handful of enormous companies, rented to you by the token. The “open source” releases are time-lagged versions of last year’s frontier, dressed up as a gift to the community and functioning, in practice, as the top of a marketing funnel. The unit economics of running a real model at home do not work for an ordinary person. Consumer hardware is artificially bottlenecked by the people who make it. The drivers don’t take advantage of the silicon you’ve already paid for. So you end up running a quantized version of a model that was already a generation behind, and if you want the real thing, you pay. And you keep paying. And what you send up, stays up.
The industry converged on this shape early, and most of its capital is committed to it. Other shapes of the technology have gone underexplored as a result.
There is another way.
What we’re building is Personal Intelligence: the full stack, from the models down to the system software that drives them, owned and built end to end. We control every layer because that’s the only way to deliver the performance and the fidelity we believe people should expect from an intelligence that lives where they are, runs on the hardware they already own, and gets better the more it knows about them without any of what it learns ever leaving their hands.
Wallace Stevens spent a career arguing that meaning is not something the world hands us, but something we have to keep making, in our own words, in our own time, against the indifference of everything that does not care whether we make it or not. The right relationship between a person and a model should preserve that. You should be able to use it without sacrificing the intelligence of the model. You should be able to use it without giving up your privacy. You should not have to choose between capability and dignity. Today, you don’t have a choice. The companies offering you the bargain pretend the bargain isn’t there.
Personal Intelligence is the alternative bargain. It gives people the ability to use AI in their daily lives without trading away the part of themselves the current model demands as payment. The first thing we are building on top of it is for memory: a place where everything you have ever lived through can live as quietly as it does in your own head, available the moment you reach for it and invisible to anyone you don’t invite in.
The things you would most want to remember are the things that make you a person, and they should belong to you all the way down.
Why does this need to be a company?
People are far more responsive to the institutions around them than we usually let ourselves admit. We move along the contours of whatever defaults the world hands us; we reach for whatever tool is closest; we trust whatever option is presented as the obvious one. The shape of a life, in aggregate, is mostly the shape of its defaults.
Right now the default is the rented intelligence in someone else’s data center, trained on the open web and increasingly on you, returning confident answers you’ve stopped checking. If you want to change what people do, you have to change what they reach for first. And the only way to change what they reach for first is to put a better thing in their hand than the one they have.
Patrick Collison once said the world is a museum of passion projects. A museum, not a marketplace; a collection of things made because someone could not stand to leave them unmade. The museum closes quietly when the default makes it slightly easier, each year, to never start the project in the first place; to let the machine choose your words; to stop noticing eventually that you’ve stopped choosing.
A company, when it works, is the legal scaffolding around a group of people who have agreed to stay in a room together until the thing exists. What we are doing, from inside that room, is building a new default that makes the existing one feel like a bargain nobody would take if they had another.
The mission is to bet against the current trends, with the work of hands. To give people back the ability to use AI in their daily lives without giving up anything they shouldn’t have to, now, while it still matters who builds it and how.
We want the people reading this essay, and the children they will raise, to live inside a world that lets them keep making meaning. Memories that are still theirs. Still ending where they end.
6OVER3 INSTITUTE was founded to make that bet. If you’d like to help, the work has begun. You can see the first of it in our short film, The Work Begins .