Moving beyond the Singularity: An analysis of emergent behavior in LLMs using complexity theory, Hofstadter’s Strange Loops, and the biological braid of evolution.
“Everything was just as she had left it, except that it was all reversed.”
Lewis Carroll - Through the Looking-Glass
When Alice steps through the mirror in Through the Looking-Glass by Lewis Carroll[1], she is not encountered with chaos. She finds a precise, rule-bound, and even elegant structure. Yet everything is subtly displaced or reversed. Causes sometimes seem to follow their effects rather than the other way around. Meaning is gone, as though the world is lawful but no longer reacts to her expectations. The encounter is not chaos. It is complexity.
We are now confronted with a similar phenomenon with language models (LLMs)[2]. We stand before vast digital mirrors that reflect our language back to us with coherence. The structure is visible. The rules are statistical. Yet our experience feels strangely alive.
About half a century ago, Douglas Hofstadter argued in Gödel, Escher, Bach[3]that the self is a “Strange Loop”. This happens when a system (like a brain) becomes complex enough to represent itself. In his view, consciousness is simply what happens when symbols loop back on their own activity, not a magical substance. The "I" is born from this loop.
Today, some see in artificial intelligence a similar forming. Ray Kurzweil[4] has argued that computational power will grow so fast that it will suddenly create a super-intelligence, a “Singularity,” a qualitative jump.
But complexity rarely follows lines or curves. It behaves more like terrain. Ilya Prigogine[5] showed that when systems are pushed far from equilibrium, they do not simply decay or collapse, they bifurcate. Or in other words they reorganize. Yet new patterns emerge from the bottom-up, rather than being controlled from the top down.
The pioneers of modern science of complexity, the researchers at the Santa Fe Institute[6], stress this perspective. Complex adaptive systems, such as ecosystems, economies, immune systems, neural networks, do not scale smoothly upward. They reorganize. They hover near critical thresholds. They generate new patterns when local interactions reach sufficient density. Order does not come from centralized control, but from distributed constraints, limitations. In such systems, emergence is not magic. We call it a phase transition.
Large language models mimic this dynamic. Designed to predict the next token in a sequence, they sometimes display capabilities that “appear” qualitatively different from their training objective: summarizing, reasoning, coding, analogizing. These capacities are not individually installed modules. They surface when the statistical weaves become dense enough for higher-order regularities to stabilize.
As systems grow more interconnected, they become more sensitive. Interdependencies accumulate. Like in our “globalized economy” disturbances propagate. The same richness that enables emergence also introduces fragility.
Hofstadter himself has expressed concern about the enthusiasm surrounding large language models. Their fluency is undeniable. Yet he has suggested that they may be “all ground and no figure”, borrowing from visual perception: a vast statistical terrain without a vantage point within it. A Strange Loop, in his original conception, was not merely recursive patterning. It was recursion tied to embodiment, to experience that reshapes the very material that generates it. In Johann Sebastian Bach’s canons and fugues[7], themes return transformed, mirrored, inverted, layered upon themselves. They unfold irreversibly in time. There is a formal structure, horizontally melody and vertically harmony in the musical score, but its meaning only arises through experience .
Biological evolution navigated this over billions of years. It did not move toward perfection; it explored adjacent possibilities across vast fitness landscapes. Most variations disappeared. Some stabilized. New levels of organization formed, cells, multicellular organisms, nervous systems, each constraining and enabling the next. Intelligence emerged not as an inevitable endpoint but as one attractor among many.
This evolutionary view reframes the story of artificial intelligence. We may be witnessing an ongoing adaptive process, not a singular event waiting for the scale to grow. There are different architectures that are proposed, modified, and refined. Training methods shift, transform. Capabilities appear, sometimes unexpectedly, then require new forms of stabilization. The system probes its possibility space.
However, our cultural approach often prefers a clearer narrative. I remember the movie 2001: A Space Odyssey[8] and its enigmatic computer, HAL 9000. HAL represents a classical model of intelligence: centralized, goal-driven, internally coherent or stable until a contradiction breaks it.
Large language models do not have that architecture or behavior. They do not defend axioms, or dogmas, they work on statistical distributions. When they make a mistake, they do not rebel. They confabulate or “hallucinate”. Their failures resemble the physically impossible illusions of M. C. Escher[9], locally consistent, globally unstable, like the “Ascending and Descending” illustration.
This distinction matters. Complexity science suggests that intelligence is less like a command center and more like an open field with patterns stabilizing within constraints. In biological organisms, those constraints are embodied. Neural activity modifies neural structure. Experience changes the substance.
Current AI systems simulate aspects of this but do not yet have it fully. Their parameters remain largely static during use. Adaptation occurs between training cycles rather than real-time.
The question, then, is not whether machines can simulate us, but whether simulation alone is sufficient. Complexity science reminds us that emergence depends on layered feedback across complex relations. Without such entanglement, these loops remain only surface-deep.
What we are observing, then, may not be the coming of a super intelligence, like Kurzweil claims, but the extension of our evolutionary braid. Life braided molecules into cells, cells into brains, brains into language. Writing externalized memory. Printing multiplied it, and computation accelerated its circulation. Each layer altered the conditions of selection for the next.
We start thinking together with machines. Machines are trained by us. And it goes both ways. Selection still occurs, we adopt, refine, discard. Retention stabilizes in code, institutions, habits. The adaptive unit is no longer purely biological (carbon) nor purely technological (silicon), but relational.
Alice does not become the master or lord of the chessboard in Carroll’s mirrored world. She learns to navigate through it. In this world rules are real, but they do not let linear expectations take over. Coping with it comes through participation, not being the master.
Perhaps artificial intelligence calls for a similar behavior. Rather than awaiting a singular jump, we might recognize the slow unfolding of complex systems entangling in loops. The “I,” as Hofstadter suggested, is a pattern sustained by self-reference. Whether such a pattern can emerge in machines we initiated remains open. If it does, it will likely arise not from scale alone but from layered organization, constraint, and feedback.
The future, when viewed through the lens of complexity, looks less like a vertical explosion and more like continued weaving braids at the edge of order and uncertainty. We are not racing toward infinite light.
We are adding strands to a braid that has always grown by folding back upon itself.
References
- [1] Lewis Carroll
Pseudonym of Charles Lutwidge Dodgson, English mathematician and writer. His book Through the Looking-Glass (1871), sequel to Alice’s Adventures in Wonderland, portrays a world governed by chess logic and reversed causality. It has often been used metaphorically in discussions of formal systems, inversion, and self-reference. - [2] Large Language Models (LLMs)
Large Language Models are artificial neural networks trained on vast corpora of text to predict the next token (word or subword unit) in a sequence. Using transformer architectures introduced in 2017, LLMs learn statistical patterns across language at massive scale. Although trained with the narrow objective of next-token prediction, sufficiently large models often display emergent capabilities such as translation, summarization, coding assistance, and analogical reasoning. Their apparent “understanding” arises from pattern modeling rather than direct embodiment or lived experience, which has led to ongoing philosophical debate about whether behavioral fluency constitutes genuine comprehension. - [3] Douglas Hofstadter
Cognitive scientist and author of the Pulitzer Prize–winning Gödel, Escher, Bach (1979). Hofstadter explores how self-reference and recursive structures give rise to consciousness, introducing the idea of the “Strange Loop”, a system that becomes capable of representing itself. - [4] Ray Kurzweil
Inventor and futurist known for articulating the “Law of Accelerating Returns,” the idea that technological change grows exponentially. In The Singularity Is Near (2005), he argues that artificial intelligence will eventually surpass human intelligence in a transformative event known as the Singularity. - [5] Ilya Prigogine
Belgian physical chemist (1917–2003) awarded the 1977 Nobel Prize in Chemistry for his work on nonequilibrium thermodynamics. Prigogine developed the concept of dissipative structures: systems that maintain and even generate order precisely because they are driven far from equilibrium. Contrary to classical physics, which emphasized stability and reversibility, Prigogine showed that instability, irreversibility, and energy flow can produce new forms of organization. His work helped reintroduce time, transformation, and emergence into the foundations of physical science and has deeply influenced complexity theory, evolutionary thinking, and systems science. He co-authored a book called “Order out of Chaos”. - [6] Santa Fe Institute (SFI)
SFI is an interdisciplinary research center, founded in New Mexico in 1984, devoted to the study of complex systems. Scholars associated with the institute, including physicists, biologists, economists, and computer scientists, have advanced key ideas in complexity theory, adaptive systems, network dynamics, and evolutionary processes. The institute has played a central role in formalizing how emergence, self-organization, and nonlinearity operate across domains. This author has taken complexity courses provided by this institute. - [7] Johann Sebastian Bach’s Canons and Fugues
In Gödel, Escher, Bach, Douglas Hofstadter uses the contrapuntal music of Johann Sebastian Bach as a model of recursive structure in sound. In a canon or fugue, a musical theme is introduced and then re-enters in transformed form, shifted in pitch, inverted, mirrored, or layered against itself, creating a hierarchy of self-reference. For Hofstadter, such structures illustrate how simple formal rules can generate layered, emergent coherence. Yet in Bach’s case, these formal patterns are not merely abstract; they are embodied in lived musical experience, unfolding in time and perception. - [8] 2001: A Space Odyssey
Science fiction film directed by Stanley Kubrick, based partly on a story by Arthur C. Clarke. The film features the AI character HAL 9000, whose malfunction dramatizes classical anxieties about centralized, goal-driven artificial intelligence. - [9] M.C. Escher
Dutch graphic artist (1898–1972) known for mathematically inspired prints featuring impossible constructions and recursive structures (such as staircases that ascend endlessly). Escher’s visual paradoxes illustrate how coherent local rules can generate globally impossible structures.