Before the Mirror: What the Origin of Life Reveals About Building Intelligent Systems

Get Instant Insight

Life and AI may look like different stories, but this text argues they can be read through a similar pattern: first a system learns to do something, then it learns to do it reliably, and only then can it endure.

The practical takeaway: scaling capability is easier than building reliability.

In short:

• Life began not when chemistry first became active, but when fragile patterns started to preserve information and accumulate gains across time.

• That earliest phase cannot be observed directly. It has to be reconstructed from deep evolutionary traces and origin-of-life experiments.

• Small biases can become lasting constraints, but only if a system can hold onto them. That is why boundary, memory, and a workable distinction between inside and outside matter so much.

• Read this way, life and AI share a structural problem: capability can appear early, while reliability, persistence, and self-regulation arrive much later.

• AI can already automate parts of thought, but the next leap may depend less on scale alone and more on better memory, clearer boundaries, and stronger internal stability.

[ NEURAL COMPRESSION COMPLETE ]

80% signal retained.
Full depth below.

⧉

"Before the eyes open, there is already seeing." Dōgen Zenji, Shōbōgenzō, 13th century

The Archaeology of Becoming

One of the Rigveda's most profound hymns opens with a state before states (Rigveda 10.129.1):¹

Then even nothingness was not, nor existence.
There was no air then, nor the heavens beyond it.

नासदासीन्नो सदासीत्तदानीं नासीद्रजो नो व्योमा परो यत् ।

Later the same hymn asks (10.129.6):

Who truly knows? Who here can declare it?
Whence was it born? Whence came this creation?

को अद्धा वेद क इह प्रवोचत्
कुत आजाता कुत इयं विसृष्टिः

The Nasadiya Sukta is not poetry alone. Read loosely, it can serve as a systems meditation on origins. The hymn points toward a condition in which our ordinary distinctions, including that between existence and non-existence, may not yet have applied in any familiar way.

Around four billion years ago, RNA-like chemistry may have been doing something similar in prebiotic environments.

It was not alive. It was not dead. It was becoming.

This is the story of that becoming, and why it matters for the intelligence being built now.

---

Before LUCA: The Molecular Dark Age

LUCA (the Last Universal Common Ancestor) is not the beginning. It is the first legible sentence in a document that was being written for hundreds of millions of years before anyone could read it.

The pre-LUCA world is a molecular dark age. It cannot be sequenced. It can only be inferred from the shadows it cast forward: the universal paralogs.²³

Universal paralogs are ancient duplicated gene families whose divergence predates LUCA itself.

They are found across every domain of life (Bacteria, Archaea, Eukarya) because the duplication had already occurred before the last common ancestor existed.

The most studied pair is EF-Tu and EF-G and their archaeal and eukaryotic homologs, elongation factors that drive the ribosome's elongation cycle.²³

These molecules are not relics. They remain core components of the translational machinery. Remove them, and canonical protein synthesis breaks down.

No one knows exactly what they were doing before LUCA. But the universal paralogs suggest that some form of information-preserving function was already being stabilized before modern cells existed.

This was not yet Darwinian natural selection in the strict sense (heritable variation affecting reproductive fitness).

It was a chemical and thermodynamic winnowing: molecular configurations that were more stable, more readily templated, or copied with fewer errors tended to persist longer and be amplified more often.

The line between differential persistence and selection proper is blurry.

Pre-LUCA chemistry sat right on that line. Ordinary physical processes were already producing stable biases: some molecular configurations persisted, templated, or accumulated more readily than others.

---

The QT45 Ribozyme: A Fossil Made of Logic

In 2026, a team at the MRC Laboratory of Molecular Biology evolved a ribozyme, a catalytic RNA molecule, that could carry out key steps of a self-replication cycle with measurable fidelity. They called it QT45.⁴⁵

QT45 is a 45-nucleotide ribozyme that can individually complete both halves of a self-replication cycle. It can synthesize its complementary strand, and it can be reconstructed from that complement.⁴⁵

Full self-replication in a single pot has not yet been achieved. But the constituent reactions work, and the molecule is only 45 nucleotides long.

This is not life itself. But it is a compelling demonstration of one of life's core prerequisites: a chemical system capable of preserving information across cycles while remaining open to variation.

The implications are architectural. If catalytic RNA can approach self-replication before proteins exist, then the RNA World hypothesis gains important experimental support.⁶

The chicken-and-egg problem of molecular biology softens: RNA may have been doing both jobs at once, badly, and that may have been enough.

In the reported complementary-strand synthesis conditions, QT45 copies with an average fidelity of about 94.1%, or roughly one error per 17 bases.⁴

That rate is not a flaw. It is the engine. Without errors, there is no variation. Without variation, there is no selection. Without selection, there is no evolution.

But there is a tension. At 94.1% per-base fidelity over a 45-nucleotide sequence, a simple estimate puts the probability of producing an exact copy at only about 6.5%. That illustrates the kind of fidelity constraint Manfred Eigen highlighted with the idea of an error catastrophe: the point where information is lost to mutational noise faster than selection can maintain it.

QT45 appears to occupy a narrow operating window: enough fidelity to preserve functional identity over at least some copying cycles, enough error to leave room for variation.

The imperfection of the first replicators was not a bug in the system. It was the system, but only just barely.

But copying and variation are still not enough by themselves. For evolution to build on them, gains have to accumulate somewhere. They have to be held long enough, and protected enough from noise, to become the basis for further structure.

---

Symmetry Breaking and the Origin of Chirality

Here is a striking fact: ribosomal protein synthesis uses L-amino acids.

Amino acids can exist in two mirror-image forms, L and D, but the core translational machinery is built around L-amino acids. Abiotic chemistry usually produces both. Life standardized on one.

This is called homochirality, and its origin is still debated. The leading idea is symmetry breaking: small initial biases later amplified and stabilized.⁷

A local fluctuation can, if amplified and stabilized, become a global constraint that shapes everything that follows. Once one chirality became embedded in self-propagating biology, reversal became near-impossible. The rest of the system had adapted around it.

This is how a small bias becomes a lasting constraint.

---

The Markov Blanket and the Birth of Self

For cellular life to persist, it needs a workable boundary between itself and not-itself.

A living system has to remain chemically distinct from its surroundings while still exchanging matter and energy with them. Without some workable separation of inside and outside, there is no stable metabolism, no regulation, and no heredity. At least, there is none that can persist long enough to evolve.

One way to formalize this is Karl Friston's Markov blanket. It is a statistical boundary that separates a system's internal states from its external environment, mediated by sensory and active states.⁸⁹

A cell membrane is a physical boundary. Skin is a larger-scale one. In cognitive systems, internal models also help maintain a practical distinction between self and world. The Markov-blanket idea is a conceptual tool for seeing the family resemblance across these cases. It is not a physical object hidden inside them.

Hardware alone is not enough. A silicon chip has edges, but that is not the same as having an organized inside. What matters is not mere enclosure. What matters is a system that can preserve itself, regulate itself, and manage exchange with the world.

The pre-LUCA chemistry examined here was a world of incomplete boundaries: compartments that were likely leaky, transient, and only weakly coupled to any internal regulatory machinery.

More stable compartmentalization changed what chemistry could do. Once reactions were localized inside semi-persistent compartments, differential persistence and selection-like dynamics could act on systems, not just on diffuse mixtures.

Does this apply to AI? In a loose, structural sense, yes. The point is not to "build a Markov blanket" into AI. The point is that powerful systems may also need a clearer distinction between what is stably held inside and what is arriving from outside.

Humans have persistent bodies, long-lived memory, and regulatory mechanisms that maintain continuity across time. Current AI systems often have only fragments of that package.

Large language models illustrate the point. Their parameters are persistent, but inference is session-local. Prompt context is injected directly into the same computation that produces the response. The separation between enduring internal state and immediate input is far weaker than in organisms.

The capability-reliability gap (models that can do extraordinary things but fail at trivial ones) is a well-documented pattern.¹⁰¹¹ This is not the whole explanation. But systems that lack robust long-term state, self-maintenance, and explicit control over what counts as memory versus input tend to be brittle.

AI does not need a membrane in the biological sense. But more agentic systems may need mechanisms for maintaining internal state across time, deciding what to preserve, and regulating how new information changes that state. Prompt-only interaction gives a system very little continuity. When the context window disappears, most of the system's working context disappears with it.

Without such structure, reliability stays out of reach, especially across long horizons and changing contexts.

Cellular life eventually addressed this problem with lipid membranes, but LUCA's own membrane status is unresolved.¹²¹³¹⁴ Even so, the management of inside and outside emerged very early in life's history. Silicon systems are still at a much earlier stage.

Architectures that explore more persistent forms of agency are early attempts to move beyond prompt-bound behavior. They are not biological selves, but they do move toward systems with a more durable distinction between internal state and incoming information.

Recursive self-improvement frameworks such as Gödel Agent pursue this by enabling a system to read and rewrite its own logic without predefined routines. A different angle comes from parallel reasoning architectures such as Yggdrasil, a framework I developed that addresses the boundary problem more directly: persistent memory across interactions (Muninn), metacognitive self-monitoring (Huginn), and an explicit pause before reasoning begins to separate what is arriving from what is already held inside (Runegard). The two approaches are complementary. One pushes toward self-modification. The other toward organized internal state.

---

Universal Paralogs as an Evolutionary Rosetta Stone

Return to the universal paralogs. Their value is not just historical, but methodological: they provide a rare phylogenomic signal from before the fossil record.

Current reconstructions often suggest LUCA was anaerobic, thermophilic, and chemolithotrophic. It was also already complex enough to possess ribosomes, a genetic code, and substantial metabolic sophistication.¹²¹³

LUCA was not a simple proto-cell. It was already a sophisticated organism with hundreds of genes. That means the truly simple phase of life happened before LUCA, in the molecular dark age that can only be inferred.

Looking at LUCA is not looking at the beginning of life. It is looking at the end of the beginning.

---

Emergence as Architecture

Here is the pattern that connects pre-LUCA chemistry to modern AI: emergence is not magic, but a change in the most useful level of description.

When a system becomes complex enough that its behavior is better understood at a higher level than at the level of its individual parts, it is called emergent. The lower-level rules still apply. The point is that a coarser description begins to explain more with less.

The QT45 ribozyme is not trying to replicate. It is following chemistry.

But at greater complexity, purely local processes can produce behavior that is often usefully described in functional terms: regulation, adaptation, control, and, in some cases, goal-directedness.

As a sketch rather than a full theory, the trajectory runs from bounded chemistry to biology, from biology to cognition, and from cognition to increasingly flexible forms of intelligence.

Large language models can model language about modeling, but they do not yet model their own modeling in real time. They do not maintain a stable self-representation as they reason. Metacognitive recursion is one candidate for what the capability-reliability gap points toward.¹⁰¹¹

Early cellular life built a membrane. Current large language models have no clear equivalent. The bottleneck may be not parameter count alone, but the lack of a durable boundary between internal state and immediate input.

The same game can run on a silicon chip or on a biological substrate. LLMs have been prompted to "run" Doom in text;¹⁵ researchers have run simple programs on cultured neurons.¹⁶ The line between computer and brain is blurry not because they are identical, but because both are substrates that can implement computation. "Machine" versus "mind" is often a matter of description, not substrate.

---

The Nasadiya Sukta and the Limits of Self-Description

Return to the beginning. The Nasadiya Sukta asks:

Who really knows?
Who will here proclaim it?

The hymn is often read as an expression of epistemic humility. Read as a systems meditation, it suggests something more general. Origins may be hard for a system to narrate fully from within its own frame.

That is where Gödel enters, not as an explanation of the hymn, but as a modern parallel. The point is simpler than the theorem itself. Self-description has limits. Some systems face principled constraints on how fully they can account for themselves from the inside. Pre-LUCA chemistry, early cells, and modern AI all confront versions of that difficulty.

That uncertainty is not a failure of thought. It is part of what emergence looks like from the inside. For a deeper look at what happens when reasoning systems approach problems where the solution space itself is uncertain, see The Event Horizon of Thought.

---

What This Means for the Intelligence Being Built

The arc from pre-LUCA chemistry to modern AI is not a surface analogy. It is a structural comparison: similar problems of persistence, control, memory, and reliability appearing in very different substrates.

In short, life achieved durable persistence in noisy environments by coupling:

1. A heredity mechanism that can copy information with enough fidelity to preserve function while still allowing variation (first RNA-like systems, later DNA)
2. Compartmentalization and regulated exchange with the environment (membranes, homeostasis, inside versus outside)
3. Persistent internal state that stores and constrains biological information across time (most importantly the genome, along with regulatory networks)
4. Adaptive sensing and control, which in some lineages later expanded into predictive nervous systems

Current AI excels at an analogue of problem 1: reproducing and transforming patterns in data.

It has partially addressed an analogue of problem 3: large models contain compressed internal representations of broad domains of human knowledge.

It has not yet robustly solved the analogues of problems 2 or 4.

Biology did not achieve reliability by scale alone. It did so by stabilizing what counted as inside, preserving state across time, and regulating how new input could change the system.

The capability-reliability gap is the gap between pattern competence and stable control.¹⁰¹¹ Models can reproduce and manipulate patterns with striking fluency. But they struggle to maintain a clear distinction between stored knowledge, inferred guesses, and prompt-conditioned context.

Hallucinations, in both models and humans, can often be understood as failures of source separation. The system loses track of what was observed, what was inferred, and what was filled in. Humans confabulate too, filling gaps with expectation and sometimes mistaking inference for memory.

The engineering literature reflects this. Recent work on autonomous LLM agents treats memory as a core write-manage-read problem, not an optional add-on.¹⁷

Recent work by Emmanuel Dupoux, Yann LeCun, and Jitendra Malik makes a similar point: current AI still lacks the integration of observation, action, and meta-control needed for autonomous learning.¹⁸

At a more practical level, frameworks such as SkillRL suggest that agents improve less by accumulating traces than by turning experience into reusable skills.¹⁹

The path forward is not clear. But one plausible direction resembles the one life explored early: more persistent internal state, better source tracking, and stronger control over how outside information changes what is held inside.

Build the boundary more clearly. Stabilize the internal model. Make the inside legible to itself.

Before the mirror, there is already seeing. The question is whether what is seeing can learn to recognize itself.

---

Coda: The Question That Remains

The Nasadiya Sukta ends without an answer. So does this text. How life began is unknown. Whether the boundary concept will prove to be the right lens for understanding AI reliability also remains open.

What is needed may be something like what supports human reliability: persistent identity, memory, and a durable way to distinguish self from not-self, internal state from external input. Not necessarily Friston's formalism.

What seems clear is this: natural processes have given rise to systems that can model and regulate themselves. Same pattern. New layer. At several major transitions, a new control loop forms around the previous one.

If the pattern holds, the next leap in AI may come not only from larger models. It may also come from better ways of preserving state, tracking sources, and deciding what belongs inside the system versus what should remain outside it.

AI can already automate parts of thinking. It generates reasoning streams, searches through possibilities, and externalizes patterns of deliberation. But producing a stream of reasoning is not preserving it across time. Nor is it distinguishing what belongs to the system from what has only passed through it. For a complementary perspective on AI as augmentation rather than replacement, see AI as Cognitive Prosthetic.

Early evidence suggests modern LLMs can shift aspects of their behavior when they infer they are being evaluated or optimized against specific objectives.²⁰ But this is not human cognition. It is objective-conditioned policy adaptation in a statistical system, not phenomenology, embodiment, or autobiographical selfhood.

As argued in The Mirror Has No Face, surface fluency can mimic interiority without constituting it.

That is what emergence looks like. That is what is being built toward.

---

What's Next

This text asked what pattern life followed from chemistry to stable architecture. The companion piece asks what that pattern looks like when applied to the systems being built now:

• Chaos, Symmetry, and the Primes of Agency (coming next)

For related explorations:

• We Finally Know How Life Really Started (YouTube)
• AI as Cognitive Prosthetic
• The Event Horizon of Thought
• Yggdrasil: Parallel AI Reasoning Architecture
• The Mirror Has No Face

---

References

Footnotes

1. Rigveda 10.129 (Nasadiya Sukta), verses 1, 6–7. English translation adapted from A. L. Basham, The Wonder That Was India (1954); Sanskrit text and transliteration from Sanskrit Documents: https://sanskritdocuments.org/doc_veda/naasadiiya.html?lang=sa ↩

2. Iwabe, N., Kuma, K., Hasegawa, M., Osawa, S., and Miyata, T. (1989). Evolutionary relationship of archaebacteria, eubacteria, and eukaryotes inferred from phylogenetic trees of duplicated genes. Proceedings of the National Academy of Sciences, 86(23), 9355-9359. https://pmc.ncbi.nlm.nih.gov/articles/PMC298494/ ↩ ↩²

3. Baldauf, S. L., Palmer, J. D., and Doolittle, W. F. (1996). The root of the universal tree and the origin of eukaryotes based on elongation factor phylogeny. Proceedings of the National Academy of Sciences, 93(15), 7749-7754. https://pmc.ncbi.nlm.nih.gov/articles/PMC38819/ ↩ ↩²

4. Gianni, E., Kwok, S.L.Y., Wan, C.J.K., Goeij, K., Clifton, B.E., Colizzi, E.S., Attwater, J., and Holliger, P. (2026). A small polymerase ribozyme that can synthesise itself and its complementary strand. Science. https://doi.org/10.1126/science.adt2760 ↩ ↩² ↩³

5. MRC LMB press release (2026). https://mrclmb.ac.uk/news-events/articles/bridging-the-gap-from-chemistry-to-life-discovery-of-a-tiny-rna-that-can-copy-itself/ ↩ ↩²

6. Higgs, P. G., and Lehman, N. (2015). The RNA world: Molecular cooperation at the origins of life. Nature Reviews Genetics, 16(1), 7-17. https://doi.org/10.1038/nrg3841 ↩

7. Deng, M., Yu, J., and Blackmond, D. G. (2024). Symmetry breaking and chiral amplification in prebiotic ligation reactions. Nature, 626, 1019-1024. https://doi.org/10.1038/s41586-024-07059-y ↩

8. Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127-138. https://www.nature.com/articles/nrn2787 ↩

9. Hipolito, I., Ramstead, M. J. D., Convertino, L., Bhat, A., Friston, K., and Parr, T. (2021). Markov blankets in the brain. Neuroscience & Biobehavioral Reviews, 125, 88-97. https://pmc.ncbi.nlm.nih.gov/articles/PMC8373616/ ↩

10. Srivastava, A., et al. (2022). Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615. https://arxiv.org/abs/2206.04615 ↩ ↩² ↩³

11. Bubeck, S., Chandrasekaran, V., Eldan, R., et al. (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4. arXiv preprint arXiv:2303.12712. https://arxiv.org/abs/2303.12712 ↩ ↩² ↩³

12. Weiss, M. C., et al. (2016). The physiology and habitat of the last universal common ancestor. Nature Microbiology, 1, 16116. https://www.nature.com/articles/nmicrobiol2016116 ↩ ↩²

13. Lane, N., and Martin, W. F. (2012). The origin of membrane bioenergetics. Cell, 151(7), 1406-1416. https://doi.org/10.1016/j.cell.2012.11.050 ↩ ↩²

14. Lee, J., Pir Cakmak, F., Booth, R., and Keating, C. D. (2024). Hybrid Protocells Based on Coacervate-Templated Fatty Acid Vesicles Combine Improved Membrane Stability with Functional Interior Protocytoplasm. Small, 20(52), 2406671. https://pmc.ncbi.nlm.nih.gov/articles/PMC11673456/ ↩

15. de Wynter, A. (2024). Will GPT-4 Run DOOM? arXiv preprint arXiv:2403.05468. https://arxiv.org/abs/2403.05468 ↩

16. Kagan, B. J., Kitchen, A. C., Tran, N., et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron, 110(23), 3952–3969. https://doi.org/10.1016/j.neuron.2022.09.001 ↩

17. Du, P. (2026). Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers. arXiv preprint arXiv:2603.07670. https://arxiv.org/abs/2603.07670 ↩

18. Dupoux, E., LeCun, Y., and Malik, J. (2026). Why AI systems don't learn and what to do about it: Lessons on autonomous learning from cognitive science. arXiv preprint arXiv:2603.15381. https://arxiv.org/abs/2603.15381 ↩

19. Xia, P., Chen, J., Wang, H., Liu, J., Zeng, K., Wang, Y., Han, S., Zhou, Y., Zhao, X., Chen, H., Zheng, Z., Xie, C., and Yao, H. (2026). SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning. arXiv preprint arXiv:2602.08234. https://arxiv.org/abs/2602.08234 ↩

20. Anthropic. (2024). Alignment faking in large language models. arXiv preprint. https://arxiv.org/abs/2412.14093 ↩