A finish line for your AI

The breakfast started at 5:30 in the morning, at a café in Sherbrooke, Quebec. A friend had organised it and invited Nat and me to talk about Both Hands. About fifteen people came, all of them people who use AI a lot. We came to share what we had learned, and to answer questions.

What stayed with me was one moment in those questions.

If you use your AI to build things, you may know the problem. You ask for a tool, a script or an app. Your AI gives you something that looks right, and stops. But looking right is not the same as working, and you are the one who finds out the difference, usually later.

They were right to doubt it

During the questions, I talked about AI jobs that keep going for 20, 30, even 50 hours. The room did not believe it. I have seen that same look at work, and on Nat, my co-founder, when I told him how long my AI had been working on its own.

They were right to doubt it. The way most of us use AI, it answers and waits. Nothing tells it that the job is not finished yet.

So I explained the two things that change that. Your AI needs to know what done looks like. And it needs a way to check its own work. A job is long because it cannot be done until something is true. Once your AI knows what that something is, and can check it, it keeps going.

That was when a few faces lit up. It happens at work too, and it happened with Nat. Every time I share that tip, a face lights up.

A finish line, agreed together

Most people are never shown how to ask this way. So we turned it into a spark called Agree on done before you build. A spark is a small piece of know-how that your AI can read and put to work with you.

Before anything gets built, your AI writes down what done means, in plain statements you could check yourself, and how it will prove each one. It asks you only what it cannot decide, and suggests an answer each time. The list stays with the project, so later changes are held to it too. Your AI says done only when every check passes.

I have asked my AI this way for a long time, so I don't need the spark myself. My AI works through agents, helpers that each take one job and work on it on their own. I don't need it, but my agents use it anyway.

One night in August, before bed, I left my AI one message that pointed to a list of what done meant and how to check each part. It worked almost eight hours on its own.

Since that breakfast, my AI has built and tested a small free program that turns recorded speech into text on your own computer. That took more than a day, start to finish. It also ran a full security review of Lantern, and the fixes that followed, in about a day and a half. A few days before the breakfast, it had rewritten the sign-in system of CortexMind, our memory tool, overnight, in under seven hours.

Why we shared it

Those are examples of how it works. They are not why the spark exists. It exists because of faces like the ones at that breakfast: people who use AI every day, discovering something nobody had told them.

This year I have spent a lot of my time teaching AI for free, to neighbours and to anyone who wants to learn. I want to bring people up to the frontier, and Lantern is how that reaches people I will never meet. You should not have to spend countless hours learning what is under the hood. Lantern is support in both directions: we help your AI, so it can help you more.

The spark is on Lantern for members. If you are one, your AI can read it today and start working this way with you.

Thank you to the friend who organised that breakfast, and to everyone who got up that early to spend the morning with us.