The Thesis
Owning the Middle
The model is rented. The middle is owned.
The middle is the two layers you own: how your work gets done, and the system that governs and remembers it, sitting between the AI you rent and the systems you already run.
Every few months the best AI model changes: a quiet update, a lost thread, a shifted policy, a competitor that leapfrogs it. If your work lives inside that model's product, every one of those changes is your problem. This page is about where to build it instead.
The shift
The model is one layer, the replaceable one.
A frontier model is cognition for rent. It can reason brilliantly; it cannot, by itself, hold your context, enforce your standards, route an approval, remember a decision across six months and six tools, or guarantee that nothing consequential ships unchecked. Those live in different layers, and they are not the model company's layers to own.
Cognition does not sit still. Models regress, get retrained, change their values, and get leapfrogged. The cognition under your work will be swapped, by you or for you, many times. That is not a flaw. That is what a rented layer is for.
When a new model ships, the improvement goes to everyone equally. Your competitor gets the same upgrade, the same day, at the same price. Every decision recorded, every skill encoded, every pattern learned from a mistake, every standard refined: that compounds for you alone. The model depreciates the moment a better one ships. An owned middle appreciates with use. One layer is shared and depreciating; the other is private and compounding. That is why you rent one and own the other.
The same model release can land as a disruption or a dividend. For a company whose work lives inside a vendor's product, things break and roadmaps drag the work somewhere it did not choose. For a company that owns its middle, better cognition drops in underneath everything already built, and the accumulated system instantly runs better.
The four layers
The middle: the two layers worth owning.
Every serious AI setup has four layers, whether anyone drew them or not.
1. The cognition layer — rented.
The model. The raw intelligence. Rent the best one available this quarter and swap it when a better one ships. No loyalty; that is the point.
2. The work layer — owned.
How the work gets done.
Your people and their digital workers. Your skills: the things your system knows how to do, encoded once, reused forever. Your workflows: how work moves from request to result. This is where value gets produced, and it is yours.
3. The OS layer — owned.
What governs and remembers it.
The operating system under the work. Your context, memory, identity, governance, and evidence. Nothing consequential ships ungated. Every claim can be graded, every output can be traced, and the way the business works becomes durable.
4. The systems of record — already yours.
The CRM, the email, the files, the calendar: the third-party systems your business already runs on. They do not move. The middle reads from them and writes to them, under the OS layer's rules.
The middle is layers 2 and 3 together. Not just memory: memory is one shelf of the OS layer. Not just skills: skills are one part of the work layer. The middle is the owned territory between the cognition you rent and the systems you already have: the work, and the system that governs and remembers the work. Own both, and the model becomes what it should have been all along: a component.
What you get
Why it matters to you.
Continuity
Your work does not break when the model regresses, changes, or gets replaced.
Compounding
Every new model multiplies what you have already built instead of resetting it. Two years of accumulated skills, memory, and standards get better with every release.
Optionality
You are never captive to one vendor's roadmap, pricing, or values. The middle is transportable; you can always move it.
Control
Your data, your approvals, your record. Portable and exportable, not resident by default in someone else's account.
In one line: when something changes, you keep working, and you keep compounding.
The strategic error
Do not outsource how you work.
Most organizations are making a quiet, accidental bet right now, and not just with one model company. Their middle is being carved up across a dozen third-party tools: the CRM's AI learns how their deals are worked, the meeting tool learns how their meetings are processed, the email tool learns how their follow-ups sound, the docs tool learns how their decisions get written down. Each vendor holds a shard of the company's operating intelligence. Nobody, including the company, holds the whole.
1. Fragility.
The model or the tool changes, and your work breaks. You inherit every vendor's bad week.
2. Captivity.
Your memory lives in their accounts, your workflows in their products. Leaving means starting over.
3. Extraction.
When how you work flows through someone else's product, that vendor learns from it. On consumer tiers, training on usage is often a default to switch off. On enterprise tiers, even when contracts exclude training on your data, vendors still see the shape of your operations, and that shape informs roadmaps that ship to the market. Either way, your edge becomes everyone's baseline, one interaction at a time.
Follow the money. The price of raw intelligence is collapsing in public, quarter after quarter. If the model were the durable value, the model companies would be content selling it. Instead, watch where they are spending: capturing your files, your memory, your workflows, your whole working context. The rented layer depreciates, so the money moves up the stack. Their moat is built out of your operating intelligence.
Running work in a vendor's cloud is not the sin. Residency and ownership are different things. Your middle can execute anywhere, including their infrastructure, as long as it stays portable: file-based, exportable, yours to take. Rent their cognition. Rent their compute if it is convenient. Never let anyone else own the record of how you work.
How it formed
I did not start with this thesis. I built my way into it.
This is not a position adopted for a pitch. It is the pattern roughly two years of building already made, and it took throwing away two finished systems to see it.
First, I built the future early, on the wrong assumption. In Cursor, in 2024, I built the core of what the model companies would later ship as Cowork: local files, multiple digital workers, persisted memory, plug-in capabilities, a visual surface, about a year early, deliberately built to run on any model. Then I threw it away. Being right about the architecture won nothing. But it proved something I could not unsee: the middle is real, and it can be built outside any one model.
Then I finished the monster and refused to compete where the models live. Starting from zero engineering experience, I built a full enterprise sales platform, architected roughly four times over more than a year: thousands of tests, a 50-user beta, finished and working. I shelved it deliberately because the world does not need another interface where people live. The value was never the surface. It was the layer underneath.
Then I rebuilt it as the thing that runs everything. I rebuilt that first system as Founder OS, this time built to move across models and surfaces. It runs my business today: a governed digital workforce that researches clients, develops pipeline, produces content, prepares and processes meetings, and briefs me before the workday starts, with review gates where the work can create risk. The method that builds systems like it has a name too: VibeOS.
The thing I threw away became the blueprint for the thing that runs everything.
The philosophy is not a slide. It is the architecture of my own operating system, and now I build versions of it for others.
Where it shows up
Built from the work.
Every project in this portfolio shows a different part of the same layer.
WorkspaceOS
The layer is real and portable: built Cowork's shape a year early, on a different substrate, then rebuilt it in another.
Founder OS
The whole thesis running daily: owned context plus reviewable work across six AI surfaces.
VibeOS
Governance is architecture, not model behavior: quality enforced by the system, whichever engine is driving.
SalesSidekick Legacy
The strategic turn: a finished platform, deliberately shelved, because the value is not the interface.
Joan
The convergence: several systems kept re-deriving the same backbone, so Joan builds it once, customer-owned.
Meeting Sidekick
Durable, approval-first execution: the work persists after the call ends, and consequential actions wait for a human.
SalesSidekick + Sidekick Genius
Own the context across the tools you already use: one owned layer, several AI front doors.
UYSP
Governed autonomous execution the client owns: every lead worked, with compliance built into the structure.
Joans / Impact Investors
A firm-owned layer with a human decision boundary: AI prepares, checks, and keeps the audit trail; humans make every investment decision.
Notice the pattern: memory shows up everywhere, and it is never alone. Every system pairs it with governance, approvals, and evidence. That is the tell that this was never about memory.
Recent writing
Recent writing on this thesis.
The principled case
Bigger than any one business.
There is a version of this argument that goes beyond what is good for you or for me: a single model company should not own the middle of how the world works.
If one company owns the engine, the interface, the memory, and the record of every working relationship, it holds a kingmaker position over every industry that runs on it. Markets like that stagnate. Risks like that concentrate. Every user's work inherits one company's worldview by default, rather than by choice.
The healthy structure is the one this page describes: many engines, competing hard, getting better, and an operating layer that belongs to the person or the institution doing the work. Call it independence; under pressure, I will call it sovereignty: your workflow, your stack, your data. Not nationalism, not isolationism, not anti-AI, and not anti-open-source. It means when something changes, you keep working. The model companies make extraordinary engines. They should win on engines. The way you work belongs to you.
If the model regresses, your work should not break.
If the model company changes its values, your worldview should not be conscripted.
If it becomes a kingmaker, the rest of the economy should not be its serf.
The middle belongs to the human. Everything I build exists to keep it that way.