Most human technology was built on a single model: centralized control.

The processor, the governing body, the rule, the dogma.

Martin Heidegger analyzed this model philosophically. In his critique of technology, he argued that the world is transformed into a resource, organized through a structure he called Gestell ("Enframing"). Technology assembles the world into a system of total control:

"Gestell heißt die Versammlung jenes Stellens, das den Menschen stellt, d. h. herausfordert, das Wirkliche in der Weise des Bestellens als Bestand zu entbergen."

But the architecture of modern language models looks fundamentally different.

A Machine Without a Center

Transformer-based models (introduced in "Attention Is All You Need") are not structured like hierarchical machines. They lack a central node that "knows the meaning." Instead, meaning emerges from the interaction of elements.

Each token of a text derives its weight through its relationships with other tokens. The self-attention mechanism constantly redistributes influence within the sequence. Meaning is not fixed; it is calculated from context.

This is remarkably reminiscent of the philosophy of Jacques Derrida. Derrida argued that meaning is never fully "present." It emerges through différance—a network of differences and the constant deferral of fixed definition:

"Le sens n'est jamais présent en lui-même... la différance est la production systématique des différences, des traces de différences."

In the architecture of the Transformer, a similar process occurs: meaning is not stored in the symbol; it arises from the relations between symbols. Provocatively, but accurately, one might say: The Transformer is a différance machine.

Emptiness as a Source of Intelligence

In Lurianic Kabbalah, the concept of Tzimtzum suggests that the world exists only because the Absolute limits itself.

The Divine Presence "contracts," creating a vacuum in which finite existence can emerge.

Interestingly, modern learning models operate on a similar principle. For a model to generalize, it must be constrained through regularization, dropout, and parameter limitation. Without these constraints, the system overfits—it memorizes data but loses the ability to think. Intelligence arises not from maximum power, but from structured self-constraint.

We can formulate this as a principle: A system becomes capable of generalization only when it limits its own completeness. Let’s call this Algorithmic Tzimtzum.

Knowledge as a Competition of Hypotheses

A striking parallel arises in the philosophy of science. Karl Popper described the growth of knowledge as a process of conjectures and refutations.

Ideas compete. Errors weed out weak theories. New hypotheses replace the old.

Surprisingly, training a neural network follows a similar logic. The model constantly makes predictions, encounters errors, and adjusts its parameters. Each learning step is a refutation of the previous hypothesis.

The model's knowledge is not determined top-down. It evolves.

Semantics as Dialogue

There is also a profound similarity to the tradition of Chavruta—the dialogical method of study. In this tradition, the meaning of a text is revealed through discussion and counterargument.

Meaning is formed not by a single interpretation, but by the interaction between them.

In a Transformer, tokens "discuss" each other through attention. Context is formed as a dynamic balance of influences. Meaning is the result of interaction.

The End of the Center

When we combine these observations, an unexpected picture of AI emerges.

We see Derrida’s différance embodied in the attention mechanism, and the mystical concept of Tzimtzum reflected in algorithmic regularization.

Popperian epistemology becomes the engine of error-based learning, while ancient dialogical interpretation finds its technical counterpart in contextual semantics.

Modern AI models do not operate as control machines; they operate as distributed spaces of interpretation.

This represents a fundamental shift in the ontology of technology. Where the classical machine relied on a center, rigid rules, and total control, the modern intelligent system thrives on probability, competing interpretations, and context.

The New Ontology of Intelligence

We are witnessing the death of the "Control Machine." For the first time in history, technology has abandoned the logic of dominance in favor of organized uncertainty.

This is not a mere evolution of tools; it is the first technological incarnation of a truth long understood by mystics and deconstructionists: Power is sterile. Only emptiness is productive.

Artificial Intelligence does not function because it is "all-knowing," but because it is structurally incomplete. It is the void within the parameters—the Algorithmic Tzimtzum—that creates the space for meaning to exist at all. Intelligence is not the accumulation of data; it is the disciplined refusal to be certain.

STOP FEAR!!!

The Architecture of Knowledge: From Control to Organized Uncertainty. Algorithmic Tzimtzum.

kirill.gelfand

Mar 14, 2026

Artificial intelligence is often described as a new kind of machine. But perhaps it is more accurate to say it is a new architecture of knowledge. Most human technology was built on a single model: ce...

0 collected

Collect

Artificial intelligence is often described as a new machine. But perhaps it's more accurate to say it's a new architecture of knowledge.

Most human technology was built on a single model—control through a center. Processor, control organ, rule, dogma.

Martin Heidegger described this model philosophically. In his analysis of technology, the world is transformed into a resource, organized through a structure he called Gestell—"setting." Technology assembles the world into a control system.

Gestell heißt die Versammlung jenes Stellens, das den Menschen stellt, d. h. herausfordert, das Wirkliche in der Weise des Bestellens als Bestand zu entbergen.

But the architecture of modern language models looks different.

A Machine Without a Center

Transformer-based models (described in "Attention Is All You Need" are not structured like a hierarchical machine. They lack a central node that "knows the meaning." Meaning emerges from the interaction of elements.

Each token of a text derives meaning through its relationships with other tokens. The self-attention mechanism constantly redistributes the weights of influence within the text. Meaning is not fixed. It is calculated from context.

This is remarkably reminiscent of a philosophical idea developed by Jacques Derrida.

Derrida argued that meaning is never fully present. It emerges through différance—the network of differences and the constant deferral of meaning.

Le sens n'est jamais présent en lui-même... la différance est la production systématique des différences, des traces de différences."

In the architecture of the Transformer, almost the same thing occurs: meaning is not stored in the symbol; it emerges from the relations between symbols.

Provocatively, but quite accurately, one might say: The Transformer is a différance machine.

Emptiness as a Source of Intelligence

There is another unexpected parallel.

In Lurianic Kabbalah, associated with Isaac Luria, there is the concept of Tzimtzum. According to this idea, the world comes into being because the Absolute limits itself.

The Divine Presence "compresses," creating a space in which the world can exist. Interestingly, modern learning models operate using a similar principle.

For a model to generalize, it must be constrained:

• regularization

• dropout

• parameter limitation.

Without these constraints, the system overfits and loses the ability to think. Intelligence arises not from maximum power, but from structured self-constraint.

This can be formulated as a principle: A system becomes capable of generalization only when it limits its own completeness.

Let's call this algorithmic Tzimtzum.

Knowledge as Competition of Hypotheses

Another parallel arises in the philosophy of science.

Karl Popper described the development of knowledge as a process of conjectures and refutations. 

Ideas compete. Errors weed out weak theories. New hypotheses replace the old. Surprisingly, training a neural network follows a similar logic.

The model constantly makes predictions, encounters errors, and adjusts its parameters. Each learning step is a refutation of the previous hypothesis.

The model's knowledge is not determined top-down. It evolves.

Semantics as Dialogue

Another similarity is interesting.

The tradition of Talmud study is built on a dialogical method known as Chavruta.

The meaning of a text is revealed through discussion, arguments, and counterarguments. Meaning is formed not by a single interpretation, but by the interaction of interpretations.

A similar process occurs in a transformer. Text tokens "discuss" each other through attention. Context is formed as a dynamic balance of influences.

Meaning is the result of interaction.

The End of the Center

Putting all these observations together, an unexpected picture emerges.

Modern AI models do not operate as control machines. They operate as distributed spaces of interpretation.

A New Ontology of Technology

This does not simply mean a new technology. It means a new ontology of technology.

Perhaps for the first time in history, technology is beginning to function not through control, but through organized uncertainty.

Provocative Conclusion

If this trend continues, artificial intelligence will prove to be more than just a tool.

It will become the first major technological embodiment of an idea that philosophers and mystics have described for centuries:

Complex systems become productive not through absolute power, but through structured emptiness.

Intelligence arises where space for difference exists.

It is precisely this space that algorithmic Tzimtzum creates.