You can offload a task, but never the learning

Every few days someone asks me which AI model they should be using. Which one’s the smartest this month, which one to switch to, whether the latest release changes everything. It’s a fair question. It’s also the wrong one — and chasing it is quietly one of the more expensive mistakes a business owner can make right now.

Picking a model is like arguing over which brand of power tool to buy before you’ve decided what you’re building. The tool matters less than you’d think. What matters is whether your business gets a little better every time it picks one up.

Here’s the part nobody selling you AI says plainly: a model on its own is just rented horsepower. Point it at nothing in particular and it’ll happily run in circles, fast, confident, and producing very little. Without someone’s direction — a goal, a standard, a reason — compute is just an expensive way to generate words.

So the real question isn’t which model. It’s: what am I building that gets smarter the longer I run it?

Two kinds of capital: human and token

To answer that, it helps to name the two things you’re actually working with. I’ll keep the labels plain.

The first is human capital. That’s everything your people carry around in their heads: what they know, the judgment they’ve earned, the relationships they hold, the knack for spotting the pattern that matters before anyone else does. It’s the veteran who looks at a job and tells you it’ll go sideways for a reason nobody ever wrote down. You can’t buy it off a shelf. You grow it, and most of the time it walks out the door at five o’clock and you hope it comes back tomorrow.

The second I’ll call token capital. When AI does work, it reads and writes in little chunks of text called tokens, and when you use somebody else’s model you pay by the token, a meter running on the wall. Most businesses only ever rent. Token capital is what you get when you stop only renting and start building something you own on top of those models: a system that captures how your business actually works, so each use makes the next one better. It’s the difference between paying for electricity and owning the machine the electricity runs.

Now the part that catches people off guard. You’d expect that as the AI gets more capable, your people matter less. The opposite is true. Human capital doesn’t shrink as token capital grows — it becomes more valuable, because human judgment is the engine that makes the machine worth anything. People set the ambitious goals. People connect dots across things that don’t obviously belong together. People build the trust that closes the deal. People know which patterns are signal and which are noise. Take that away and you’re back to compute running in circles, tireless and busy and pointed at nothing.

Build the loop, not the shopping list

So if buying a model isn’t the move, what is? Building the loop.

Picture a cycle that turns every day of work into something you keep. Your people do the work and make the calls. The system captures how they did it and why. Next time a similar job lands, that captured know-how is right there, so the work starts further down the road instead of from scratch. Do that enough times and the business stops repeating itself. It compounds.

The move, in one sentence, is to turn your workflows, your know-how, and your hard-won judgment into systems that get better every time you use them. Here’s the line I keep coming back to:

You can offload a task, or even a job — but you can never offload your learning.

You can hand a task to an assistant, a contractor, or an AI. You can hand off a whole role. But the learning, the part where the work teaches you something and you come out sharper, has to land somewhere you own, or it evaporates. For most businesses today, it evaporates: the knowledge lives in one person’s head and leaves when they do, and the AI forgets everything the second you close the window. A learning loop is just the decision to stop letting that happen, to catch the learning and keep it.

That’s the real opportunity. Not the smartest model. The loop you build on top of whatever model you’re using, where your people and your owned AI capability get better together.

The sovereignty test

There’s a catch worth being honest about, and it’s the whole reason this matters for ownership.

If the place your learning lands belongs to someone else, locked inside a vendor’s product, welded to one company’s model, then you don’t actually own the loop. You’re renting it, and the landlord can raise the rent, change the terms, or shut it off. The knowledge your business spent years earning becomes a hostage.

So here’s the test I’d hold any of this up against:

A model, even a great one, is a generalist. It knows a little about everything and nothing about you. The expertise worth real money is the veteran: the accumulated sense of how your business specifically does things. The architecture that matters keeps those two apart. The model stays swappable, a commodity you upgrade or replace as the market moves, while the veteran, your institutional knowledge, stays yours, under your roof, no matter whose model you’re renting this year.

Build it that way and a better model dropping next month is good news, not a fire drill. You swap the engine and keep everything you’ve learned. That’s what control looks like in the era ahead.

How the loop actually gets built

In practice the loop comes together in three layers. I’ll be straight about which are real today and which are where this is headed. Over-claiming is how the AI industry burned through everyone’s trust, and I’m not interested in adding to it.

A queryable knowledge base. Your institutional memory, made searchable. Every decision, lesson, and bit of hard-won context written down once, in a form the AI can pull up instantly, so it stops re-asking what it should already know, and every token you spend goes further. This part is live and running this desk today.
Private evals. An eval is just a test. A private eval is your test, built from your business’s real outcomes, so when a new model shows up you can measure whether it’s actually better at your work, not better at some public leaderboard that has nothing to do with you. This is the direction it’s built toward, not something I’m claiming as finished.
A private training environment. The far end of the loop: a practice gym built from your real work (the technical name is a private reinforcement learning environment) where a model gets measurably sharper at your specific tasks by learning from how your business actually operates. Also where this is headed, not a thing that’s live today.

The honest status: the memory layer is real and proven. The evals and the training are the road it’s on. I’d rather tell you that than sell you a finished picture that doesn’t exist yet.

This isn’t hypothetical. It’s running this desk

That first layer isn’t a slide. It’s the thing running this very desk: a private memory system the agents read from before they start work and write back to as they learn. (We’ve been calling it Satori-Kura. The name’s a placeholder, but the system is real.) It’s worth walking through what it actually does, because each piece maps to something an owner cares about.

Put those together and here’s the throughline any owner should care about: this turns one person’s domain knowledge into something the business owns, knowledge that can always be resourced. It survives the employee who leaves and the model you swap out. The veteran’s expertise stops walking out the door at five o’clock. It stays.

The owner’s move

So, to bring it home.

The biggest asset most businesses are sitting on isn’t on any balance sheet yet: the accumulated knowledge of how they specifically win. Right now it’s scattered across people’s heads and lost a little every time someone leaves. Token capital is what you get when you decide to capture it, to build a system that turns that knowledge into something the company owns and that gets sharper with use.

Your people are the engine. The model is rented horsepower. The loop is what you build so the two compound instead of evaporate. And the test of whether you actually own any of it stays simple: can you change the model without losing what your business knows?

The owner’s move, then, isn’t to go shopping for the smartest model. It’s to start building the loop.

If you run your own world and you’ve been wondering where to start with all of this, that’s the start. Not the model. The loop.