This one came to me sideways, the way the useful ones usually do. I was moving a project between two AI coding tools this week - something I do constantly, because the models trade places at the top every few weeks and I follow the better one - and I had the oddest flash of recognition. The screen looked like 1996. Not visually. Structurally. I was standing inside America Online again.
The Keyword
If you weren't there: AOL in the mid-nineties was not a way onto the internet so much as a replacement for it. It had its own mail, its own chat, its own news, its own shopping, its own everything, and a business that wanted to exist online built a section inside AOL. Television ads ended with "AOL Keyword: Pepsi." Magazines ran a keyword next to the web address, and plenty of companies ran the keyword alone. You built your storefront in somebody else's mall because that's where the people were, and for a while it worked.
Then the mall emptied. Everyone who had put the asset inside the walls had to start over on the open web. Everyone who had treated AOL as a doorway and kept the real thing on their own domain just kept going.
We've run this cycle at least three more times since. In 2008 the advice was get a Facebook page for your business, and a lot of small businesses let the page become the business, right up until the algorithm changed what a page was worth. Then it was Instagram, then TikTok, and each time a platform offered you a free, beautiful, well-trafficked section inside its walls, and each time the platform ended up owning the audience and the archive. Society moves through these in waves, and the wave always feels permanent from the inside.
The current wave is: get your data inside the model. Put your documents in a project. Upload your knowledge base. Turn on memory. Get your business into ChatGPT, into Claude, into Gemini, into Grok. Build your magazine section inside the AI.
Two Layers
Here's the distinction that I think separates the people who will be fine from the people who will be starting over in 2031.
An AI platform is two different things wearing one interface. There is a computation layer - the model, the brain, the thing that reasons and writes and codes. And there is a data layer - your documents, your history, your structured knowledge, your decisions, the accumulated record of what you know and how you work. The interface blurs them on purpose, because a chat window that remembers everything feels like one magical object. It isn't. It's a brain borrowed by the hour sitting on top of a library you've been quietly handing over.
And those two layers have completely different economics. The brain is a commodity that keeps getting cheaper and better. I publish monthly rankings of these models, and the lead has changed hands repeatedly this year alone. Switching brains costs me an afternoon. Sometimes it costs me nothing, because the tools are built to be pointed at the same files.
The library is the opposite. It is slow to build, impossible to regenerate, and it compounds. Twenty-four years of Hawaii knowledge does not come back if the house it lives in changes the rules. That asymmetry is the whole argument: the layer that's easy to replace is the one they're selling you, and the layer that's impossible to replace is the one they're asking you to store with them.
The Sticky Layer
I want to be fair, because this is not a conspiracy. Projects, memory, custom instructions, and knowledge uploads are genuinely useful. I use them. But look at which layer every one of those features lives in. None of them makes the brain stickier, because the brain can't be made sticky - a better one ships next month. All of them make the library stickier. They are switching costs, and the vendors know exactly which layer holds the switching cost, the same way AOL knew your email address was worth more than your subscription.
That's not a reason to refuse the features. It's a reason to be clear about what they are. Anything you put inside a platform's memory should be a copy. The original lives somewhere you control, in a format any brain can read.
How the Shop Already Runs
I didn't arrive at this as a theory. I noticed it because it's already how we work, and the theory is just me catching up to the practice.
The footer of my logs index says these posts are coded and written with help from Claude Code and Codex. Both. Interchangeably. The posts themselves are plain HTML files in a repository I own, and when one model falls behind, I point the other at the same folder and nothing about the library changes. On the HawaiiGuide side it's the same arrangement at a larger scale: the ground truth - the ocean data, the drive times, the reservation rules, the hurricane-season alert history, the structured briefs we publish for other people's agents - lives in our repositories and our database, in formats we chose. The model that writes code against it has changed several times this year. The data has never moved.
When hurricane season tested the live-conditions system this fall, the brain was whichever model was best that week. The thing that actually held was the library: the vetted sources, the accumulated judgment about what a visitor needs, the record of what the islands were doing day by day. That's the asset. The brain is a contractor.
The Rule
So here is how I'd articulate it, for anyone who's serious about building with these tools rather than just using them:
Create the tool you need in your own house, using whichever brain you currently value. Don't let the data live in a house that isn't yours.
In practice that means a few unglamorous habits. Keep the canonical copy of everything that matters in storage you control, whether that's a repository, a database, or a folder of plain files. Prefer boring, open formats over anything proprietary, because boring formats are the ones every future model will read. Treat a platform's memory and project features as a cache that can be rebuilt from your originals, never as the originals. And when you choose a tool, ask one question first: if this company disappeared tomorrow, or tripled its price, or changed its terms, what would I lose? If the answer is "an afternoon," you've chosen well. If the answer is "years," you've built a keyword.
The brains are going to keep getting better and cheaper, and you should keep swapping them without sentiment. The library is the only part of this that's actually yours. Swap the brain. Keep the library.
