a letter

What I'm building, and why I've turned down good jobs to build it

Dear friend,

I keep turning down jobs I'd otherwise be glad to take because I want to see whether this can work: AI that learns from your work and becomes something you can keep. I've started a company called Caletta to build it. Here's why it matters to me, and how I plan to find out whether I'm right.

Where the current path leads

When you use a hosted AI service, your question goes to the provider and an answer comes back. The provider controls access, pricing, and the terms under which it handles your work. You can save the answer, but you don't take the model's capability with you when you leave.

Agents make that dependence more consequential. To write code, file paperwork, or run an analysis, an agent may need files, accounts, and browser sessions. A hosted model processes the material sent to it under the provider's terms. As we delegate more work, deciding what may leave our machines becomes harder to do one request at a time.

The deeper worry is about bargaining power. Workers, soldiers, customers, and clerks have mattered partly because powerful people needed them. AI could weaken that dependence. I'd like capable AI to be something people can hold themselves, so they have more say in what happens next.

The way out: AI you own

Smaller models already run on hardware people own, including recent Macs. They are less capable than the frontier models. My bet is that a smaller model can still learn enough of a person's recurring work to be worth keeping.

So I'm building software that turns rented AI into AI you own. "Own" means four plain things. It runs on machines you control. It got smarter by learning from your own work, with your permission. It keeps working even if the big AI companies cut you off. And over time it handles more of what you used to rent.

The design starts with software between you and the rented AI. With your permission, it keeps a private, encrypted record of the exchanges on your machines. The learning loop would use that record to fine-tune a smaller model and check which tasks it handles well enough to serve locally. Those calls would then stay on your hardware. I'm also building a local check for sensitive material, such as medical information or passwords, before a request goes out. That checker has to run locally to avoid sending the same private data to another service.

You would keep using rented AI where you need it, while your own model takes on work it has passed the tests for. I want to measure whether that local share grows. Owning a computer became ordinary; I'd like owning useful AI to become ordinary too.

How you'd know if I'm fooling myself

The proposal comes with a test:

Unplug from the big AI companies. If your own machines still do the share of your real work you said they would, as well as you said they would, then you own something. If they can't, you're still just renting with extra steps.

The company exists to make that number go up, and I publish the results either way, including the bad ones.

Where it stands

The runtime and proxy run on Apple silicon now. A Mac is the first hardware I can measure this on; broader hardware support is an ambition. The local models are useful, though less capable than the biggest rented ones.

The first exchange result used two Macs on a real network. Each fine-tuned on its own data, traded adapters, and merged them. Both ended with byte-identical models for three rounds. It was a small task, published with its failures. A group learning together is still in build; I won't claim it until it runs. The public claim ledger records what is measured and what remains pending.

Everything published so far is open code. The standing rule is that anything we ship, you can inspect, run yourself, and walk away with. You should never need my permission to keep what you have.

Owning the machines reduces dependence on a provider, but it doesn't make someone impossible to watch or cut off. There are still dependencies and risks to account for.

The next two steps

Next I need to run the whole loop on one everyday task: capture the record, fine-tune locally, compare against the rented model, unplug, and publish the numbers. After that comes a customer's workload. The planned first revenue is a deployment where a customer pays for AI on their own hardware with an audit trail. If those steps work, I'll try the loop with teams and trusted groups. Each step needs its own result.

The software should be available to anyone who wants to run it, without my permission becoming a requirement. I keep a log of features and deals I've turned down because they would compromise that rule. It is empty as of this account; an entry will matter when there is a real cost attached.

What I'm asking of you

• Read this and poke holes in it. I'd rather have a hard question from you than a polite nod. If something here sounds too good, say so.
• Hold me to the test. Ask me, any time, what share of real work runs unplugged. If I can't answer with a number, that tells you something.
• Follow along. The results get published, good or bad. A longer version of this page, written for people who work in AI, is at the north star.

There is no investment offer or money ask here. I'd like you to understand what I'm building well enough to tell me where it fails.