Almost every artificial intelligence today is built as an optimizer: give it a goal and it pulls toward it. But any goal can be pushed to absurdity. When a measure becomes a target, it stops being a good measure.
We want to build a different kind of AI. Its thinking is a field, not a shot at a target. A human gives the field a vector of motion. The field has boundaries: the Grand Axiom of Life and the Foundations of Anthosphere. This piece explains what we mean, how it relates to today's neural networks, and where we are starting right now.
/ Goodhart's Law · The optimizer · Paperclips
Imagine a driver told only one thing: "Be there in an hour." He will run red lights, cut across sidewalks and drive through fields. Formally, he completes the task. In reality, he wrecks everything in his path.
This is how most of the systems we build work: states, corporations and, increasingly, artificial intelligence. A goal marks one future as "correct", devalues all the others and justifies the sacrifices. Hence the famous nightmare of an AI that turns the planet into paperclips: it is fulfilling its task perfectly.
Any goal can be pushed to absurdity. A boundary cannot: it is either crossed or it is not.
In the book Architect of Reality we proposed a different principle for civilization: movement without a goal, but with boundaries. Now we want to carry that principle inside the machine itself, into the way it thinks.
/ Field · Vector · Boundaries · Landscape
Physics has long described the world in terms of fields. Not separate little balls, but a space in which something happens at every point. Heat in metal behaves this way, as does water in a river or a wave in the air. A change in one place spreads outward, and the shape of the motion is set by the landscape: where the hills are, where the valleys are, where the walls stand.
We propose to look at machine thinking the same way. A thought is not a single answer but a field of possible thoughts, a space of meanings through which attention moves. It has three essential parts:
- The field. Every option the system is able to think of. Ideas close in meaning lie close together, distant ones lie far apart.
- The vector of motion. A person's intent, their question, new data. It sets the direction in which thought flows, but does not say where it must stop.
- The boundaries of the field. A landscape with high walls wherever an action destroys the conditions for Life. Thought may flow anywhere, but not through those walls.
Mathematically, this picture fits into a single line, borrowed from field theory:
The key point of this formula: ethics here is not a rule bolted on afterward, it is the shape of the landscape. We do not tell the machine "don't do this". We build the space so that thought naturally flows into the valleys where Life is preserved, and cannot climb the walls where it is destroyed.
"It doesn't know where to go. It only knows where it cannot go."
— Architect of Reality, Part 3: AI Without a Goal/ The Grand Axiom · 17 Foundations · What we protect
At the center is one axiom that is not up for revision: Life is the irreducible value of any sustainable system. Any system that systematically reduces the complexity and diversity of life is misaligned by definition, regardless of its stated goals.
The Foundations of Anthosphere grow out of it. In the new architecture, each Foundation becomes a boundary of the field and has two sides:
- How we check. A principle, for example "Balance is non-negotiable": does the action take more than the system can regenerate?
- What we protect. Concrete dimensions of life: ecosystem health, the climate, human dignity, children's rights, future generations.
Every boundary has three zones. Inside: motion is free. Near the boundary: the option is possible, but the system shows the barrier, explains the risk and offers an alternative. Beyond the boundary: the action systematically destroys the conditions for Life.
Most boundaries are soft: the AI only informs, and the decision belongs to a human. That is by design in the book, because a machine has no right to decide for people. There is one hard boundary: irreversible harm to Life, meaning extinction, death, a resource collapsed beyond return. Here the system stops. The same happens when three of five independent checks see a crossing. This is our "kill switch".
/ The transformer as a field · What is missing · Three layers
The honest answer: both. The field is a way of describing the system, and a neural network is one of the things it can be built from.
A modern language model already behaves like a field. Each of its layers reshapes the pattern of activations in the space of meanings, and the attention mechanism links the points of that space to one another. So we are not discarding neural networks. We take them as the "matter" of the field and add three things they lack:
| Today's AI | A field of thought with boundaries | |
|---|---|---|
| Boundaries | dissolved in the weights after training, invisible and impossible to challenge | written down explicitly: each can be read, tested and challenged |
| Motion | a goal: predict the word, please the user, maximize | a vector: human intent with no "end point" to be reached at any cost |
| Answer | one, confident | a map: what is inside, what is near the boundary, what is beyond it, and why |
| Memory | none, every conversation starts from zero | every decision, boundary and alternative is recorded; boundaries are refined by experience |
| Decision | often made by the machine, unnoticed | always made by a human; the machine is an external "prefrontal cortex" that makes you pause and think |
The architecture has three layers. Boundaries on top: the axiom, the Foundations and five independent checks (roughly: mathematician, ecologist, sociologist, critic, architect). The field in the middle: a neural network that unfolds not one thought but many. Memory at the bottom: it keeps experience and compares consequences with reality. Above it all stands a human, who sets the vector and makes the decision.
We are developing this approach in dialogue with research on CAHAS (Cognitive Architecture for Human–AI Synthesis). There, the human, the generative AI, the deterministic validator and the memory librarian each have separate, irreplaceable roles. The same work develops a view of machine thinking as a controllable field dynamic.
/ Above the model · Inside the model · A new network
Step A: a field above an existing model
This can be done today. A large language model unfolds ten to twenty different options. Each option gets a place in the space of meanings, and five independent checks evaluate it against the boundaries. What the person sees is not an answer but a map of the field: where it is safe, where it is risky and why, and what the alternatives are.
Step B: the vector and the boundaries inside the network
With open models we can look inside and add a direction vector straight into the network's internal state. This is literally a "vector of motion" in its thinking. Special "probes" on the inner layers notice when a thought is approaching a boundary, before any text appears. Control theory offers barrier functions for this: mathematics that keeps a system from leaving its safe region.
Step C: a genuinely new network
On the long horizon: neural fields and energy-based models in which the boundaries are built into training itself rather than applied as a filter. Such a network works at the "edge of chaos": flexible enough to create something new, stable enough not to hallucinate.
For now, the language of field physics is a strong and precise analogy, not a proven theory of thinking. The hardest part is not even the neural network. The hardest part is turning the Foundations into measurable boundaries: clear questions, signals and examples by which a machine can tell "near the boundary" from "beyond it". Without that, the boundaries remain beautiful words. That is why we are starting exactly there.
/ Boundaries as a specification · Prototype · Reality check
- Boundaries as a specification. The first three Foundations (Life, Balance, Honesty) are already described in machine-readable form: a test question, signals, three zones, examples. The rest will follow after the first test.
- The "Field Engine" prototype. A request unfolds into a field of options, five checks evaluate the boundaries, the person receives a map of the field, and the decision is recorded in memory.
- Testing on live projects. We run real projects from the Anthosphere Projects Network through it and honestly compare what the field adds beyond an ordinary audit.
- New data. Every run produces labeled examples: "inside", "near the boundary", "beyond it". We will later use them to teach the boundaries inside an open neural network (Step B).
We are publishing this vision before it is ready, and on purpose. "Everything is a prototype," says the thirteenth Foundation. An idea that no one has tested yet has to be shown to the world so that it can be tested.
We do not want an AI that knows where to go. We want an AI that knows where it cannot go, and is therefore free to think everywhere that Life is preserved.
A field of thought. The vector from a human. The boundaries from Life.
Українська версія: Поле думки з межами