Eunoia is not a model.
It is the objective. Omega, Zeta and Sigma are three separate attempts at parts of it, built on different assumptions, and Eunoia is what they add up to if they work.
Why it is not one system.
The pieces missing from current systems are not variations of one problem. Abstract reasoning that improves with experience, a state that evolves on its own, and a grounded model of the world are three different gaps, and we do not think one architecture gets all three by being scaled up.
So they are being attacked separately, by systems that make different assumptions and can fail independently. If one of them dies, it should die on its own — not take the others with it.
A wooden log and an iron bracelet.
Suppose someone hands you an iron bracelet and a wooden log, and asks you to make the bracelet float.
The bracelet will not float on its own. Iron is denser than water and nothing you do to the bracelet changes that. You cannot polish it into floating. You cannot make it thinner until it floats. You cannot throw it harder and hope.
You fit it onto the log.
The trick is that flotation was never a property of the bracelet. It is a property of the whole system the bracelet is part of. What matters is displacement, not strength — how much water the thing pushes aside, not how hard any one part of it is working.
We think general intelligence behaves the same way. It is the bracelet: dense, valuable, and sinking on its own.
The current answer is a toothpick.
Almost every serious attempt at general intelligence today is an attempt to float the bracelet on a language model, and to make up the shortfall with force. More parameters, more data, more compute — brute strength applied to a shape that does not displace enough water.
This is not a criticism of language models. A toothpick is not badly made. It is extremely good at being a toothpick, and we build on one ourselves. The problem is the job, not the object.
Why you cannot just use more toothpicks.
The obvious objection: lash enough toothpicks together and you have a raft. If the answer is more displacement, why is scale not the answer?
Because the raft is not available. You cannot stack language models inside a single model. You can route between them, ensemble them, mix them — and what you get back is still a language-model-shaped system with language-model-shaped limits. Generality does not appear because you placed more of the same thing next to itself.
A log is not a quantity of wood. It is wood with structure — grain, continuity, a shape that was grown rather than assembled. That distinction is the whole of our position: architecture, not amount.
So we are building the log.
A language model is part of it. So is a model of the world, so is a state that persists and evolves, so is a way of reasoning that gets cheaper with experience instead of starting over. None of those is sufficient. Our claim is that together they displace enough water.
That is why there are several programmes here rather than one, why they make different assumptions, and why each is allowed to fail without ending the others. We are not building a bigger toothpick.
This is an argument, not a result. Nothing on this page has been demonstrated, and the honest version of our position is that we think this is right and intend to find out.
The programmes.
Eunoia Omega
Reasoning that should get cheaper the more of it you do — at a fraction of the scale the field currently assumes is necessary.
Eunoia Zeta
Most models process from input to output. Zeta is built around a state that keeps changing on its own, and treats the settling of that state as the thinking.
Eunoia Sigma
A system that predicts what happens next in the world rather than what word comes next in a sentence.
Nothing here is finished, and we would rather say so.
No model on this site has published results. Every programme states in advance what would count as failure, and each carries a threshold at which we stop rather than continue on hope.
We would rather be a lab that publishes a negative result than one that publishes a number nobody can check.
What would count as progress.
- Cost that falls with experience rather than staying flat across problems.
- Continuity that survives between sessions instead of resetting each time.
- Plans that can be rejected internally before anything acts on them.
- Behaviour that holds across a long task instead of degrading with length.
What Eunoia is not
- It is not a product, and there is nothing to request access to.
- It is not a single model, a single architecture, or a single paper.
- It is not a timeline. We plan around generations rather than releases.