A body with a landlord
The manifesto opens by saying what it is: the codex's normative source, and where the mechanism conflicts with the statement, the statement is what survives. So the tenets are not decoration on top of the engineering. They are the thing the engineering answers to.
Two of them are being tested in public right now, by companies that have never heard of this project.
IV. A brain with a landlord
Centralized intelligence is a brain with a landlord. Decentralized intelligence is a brain with a billion shareholders.
Written about models. It reads differently once the brain has a body.
Tesla ended car production on the Fremont Model S/X line in early May, spent four months converting it to humanoid assembly, and has been running units inside its own plants since. Estimates in the trade press put the internal fleet somewhere near a thousand. Nobody outside the company knows, because Tesla has never published a count.
Those machines have a landlord. Not metaphorically. They are owned by one firm, deployed on floors that firm owns, and every episode they record improves exactly one model.
XIII. True safety is born in sunlight
We reject the idea that safety requires secrecy. True safety is born in sunlight. Open weights, open data, open governance. Transparency is not a risk. It is the foundation.
That tenet makes a prediction, and this summer ran the experiment.
NVIDIA launched Cosmos 3 on 31 May: an omnimodal world foundation model over text, image, video, ambient sound and action. Super runs 64B in a datacentre, Nano runs 16B on a workstation. The models are on Hugging Face. NVIDIA claimed first place among open models on Physics-IQ, PAI-Bench and R-Bench for world generation, and the lead on RoboLab and RoboArena for action policy.
Then Spirit AI took RoboArena off them, 1,924 against Cosmos3-Nano-Policy at 1,881, on DROID robots in real environments rather than in simulation.
Read that as the tenet reads it. NVIDIA put the weights in sunlight, someone smaller pulled them down, ran the comparison, and beat them where anyone can check. The claim was contestable, so it got contested, and the field is further along than it was.
Now apply the same test to the humanoid fleet. Tesla's Q1 production report counts 408,386 vehicles produced, 358,023 delivered, and 8.8 GWh of energy storage deployed. It does not mention Optimus once. A company that counts cars to the individual unit publishes, for its robot programme, no production figure, no uptime, no task volumes, no external sales.
This is not hypocrisy on Tesla's part. It is consistency. If the asset is the data, disclosing the fleet discloses the asset. Tenet XIII is not describing an oversight. It is describing an incentive, and the incentive is behaving exactly as the tenet says it does.
XII. An intelligence trained by a monopoly will reflect a monopoly
Alignment cannot be delegated to the few. An intelligence trained by a monopoly will reflect a monopoly. An intelligence trained by the world will reflect the world.
This is the tenet with the most to say about robots, and it is not primarily an argument about fairness.
It is an argument about competence.
Suppose the fleet works. Suppose a thousand machines spend two years on automotive assembly floors and the resulting policy is superb. What it will be superb at is automotive assembly floors. It will have seen a particular lighting regime, a particular set of tools, a particular tolerance for failure, a particular floor surface, and almost nothing else. It will not have been in a kitchen, a clinic, a greenhouse, a workshop, a classroom or anyone's home.
A corpus harvested inside one firm's buildings produces a machine fluent in one firm's buildings. That is not a flaw in Tesla's execution. It is the shape of the input, and no amount of capital changes it, because the constraint is not money. It is the number of distinct places you are standing.
An intelligence trained by the world reflects the world. For bodies, "the world" means rooms that nobody's shareholders own.
I. Tenancy, with hands
The moment it becomes private property, humanity becomes a tenant in its own mind.
A post here last month argued that physical interaction data is a distribution problem rather than a capital problem. Optimus is the sharpest objection to that, and it deserves a straight answer.
If you manufacture the bodies, capital does buy presence. Tesla does not have to persuade anyone to host a robot. It builds the machines and installs them on its own floor, and by the reported account the early units are there to learn rather than to be useful. That is a vertically integrated solution to the bottleneck we called hard, and it functions.
The claim was never that one actor cannot accumulate embodied data. It is Tenet I: when one does, the corpus becomes private property, and everyone who later needs a machine that can act becomes a tenant of whoever holds it. Tenet XI puts the failure mode in one line. A brain owned by one company can be shut down by one boardroom. A body fleet is the same sentence with a physical corpus attached.
VIII and IX. What the mechanism is for
Every token is a GPU hour. Every holder is a co-owner of the future being computed.
We do not have shareholders. We have stewards. We do not have customers. We have participants.
Set Cosmos 3 and Optimus side by side and the same thing is missing from both. NVIDIA published a model and collects no episodes. Tesla collects episodes and publishes no model. One gives away the artefact and keeps no data. The other keeps the data and gives away no artefact.
Neither pays the person who produced an episode, because neither has any reason to. Tesla owns the floor. NVIDIA is not standing on one.
That payment is the mechanism the tenets are asking for, and the codex specifies it. Phase 4 activates manipulation and locomotion, and its scope is a multi-embodiment registry, an embodied data contribution pipeline, and a per-episode reward. The hourglass is what makes a contributed episode worth anything to anyone else: embodiment is one of seven surfaces sharing a single learned latent, so an episode from one robot in one room is not stranded inside a robot-specific model. P2 writes Tenet IV into law, that representations are owned by the network rather than by any model that uses them.
A participant with one robot in one kitchen is the unit this design is built around. That is Tenet IX in hardware.
II and XIX. The tone, and the reason for optimism
A handful of corporations are racing to own the mind of the future. We reject this future, not with anger, but with architecture.
Which is why none of the above is an attack. Tesla is executing a coherent strategy and NVIDIA published something genuinely useful. The argument is not that these companies are behaving badly. It is that the architecture decides the outcome, and both architectures hand the result to a balance sheet.
The future will not be written by those with the most compute. It will be written by those with the most conviction. Conviction scales. Conviction compounds. Conviction is the only resource that cannot be monopolized.
And there is already a small proof of that on the board. Spirit AI has more conviction than compute and it took the top of RoboArena from NVIDIA anyway, in the open, because NVIDIA made it possible to try.
Tenet XX is the short version. Intelligence belongs to everyone, or it belongs to no one worth trusting. The bodies are the next place that sentence gets tested.