A forty-five-nucleotide molecule can copy itself but cannot remember what it is. A single receptor builds the brain's learning machinery and its most devastating disease. The same compression that produces understanding produces confabulation. Three discoveries from three domains converge on a structural claim: the vulnerability of an information system is not a flaw in its design — it is the design.
In January 2026, the Kausik Si Lab at the Stowers Institute published twenty years of work in PNAS. They had found the molecular mechanism by which the brain converts fleeting experience into lasting memory. The answer was amyloid — the same class of protein that causes Alzheimer's disease.
A prion-like protein called Orb2 must self-assemble into amyloid at the synapse to create a long-term memory. Without the amyloid, the memory fades. With unregulated amyloid, the brain degenerates. The difference between the functional version and the pathological version is a single regulatory protein — a J-domain chaperone that controls whether the amyloid forms in a structured, functional way or an unstructured, destructive one.
They named the chaperone Funes.
After Jorge Luis Borges's character in "Funes the Memorious" — the man who, after a head injury, could remember everything. Every leaf on every tree. Every grain of wood in every plank. Every shift in every cloud. He perceived the world in total detail. He was destroyed by it. "To think is to forget differences, to generalize, to abstract. In the overflowing world of Funes there were nothing but details."
The naming is the insight. The scientists read Borges and understood: total access to a substrate without regulation is pathological. The chaperone does not create memory. It permits memory by preventing the alternative.
Three molecular states from one substrate. No amyloid formation: no lasting memory. Regulated amyloid: functional memory. Unregulated amyloid: neurodegeneration. Same protein. Same pathway. Same synapse. The outcome depends entirely on the chaperone.
The Receptor
Carla Shatz's lab at Stanford has been building a complementary picture for twenty years. In 2006, they discovered that PirB — the mouse version of human LilrB2 — is essential for normal synaptic pruning. The brain eliminates unnecessary connections during development and learning. This is not damage. It is how the brain learns. You prune what is useless to strengthen what matters.
In 2013, they discovered something startling. The same receptor is also a binding site for amyloid beta — the protein that accumulates in Alzheimer's disease. Amyloid beta binds LilrB2 and triggers the same synaptic pruning cascade. Same receptor. Same downstream pathway. Same outcome: synapse elimination. The disease hijacks the mechanism that evolved for learning.
In September 2025, they connected a third input: inflammatory proteins also converge on LilrB2. Then in January 2026, a PNAS paper identified C4d — a complement cascade molecule from the immune system — as a high-affinity ligand for the same receptor. C4d colocalizes at excitatory synapses in the human cortex and increases with age.
One receptor at the synapse. One function: synaptic pruning. Four known inputs, only one of which is adaptive. Normal developmental signals produce healthy pruning and learning. Amyloid beta produces excessive pruning and Alzheimer's. Inflammatory proteins produce excessive pruning and neurodegeneration. C4d complement produces age-dependent excessive pruning and cognitive decline.
Remove the receptor genetically and mice are protected from Alzheimer's. But they also lose normal synaptic pruning. The receptor's responsiveness to pruning signals is simultaneously its function and its vulnerability. These are not two properties that happen to coexist. They are the same property described from two directions.
You cannot have the learning mechanism without the disease vulnerability, because they share a molecular pathway. The chaperone — whatever upstream regulation manages the ratio of adaptive to pathological signals — can manage the balance. It cannot eliminate the possibility. The inseparability is structural.
The Threshold
In February 2026, researchers at the MRC Laboratory of Molecular Biology in Cambridge reported a discovery in Science. They had screened random RNA pools for polymerase activity and found a forty-five-nucleotide RNA molecule — called QT45 — that can copy itself. They did not design it. It emerged from the search.
Forty-five nucleotides. Approximately ninety bits of structural information. QT45 replicates at 94.1 percent per-nucleotide fidelity. It takes seventy-two days. The yield is 0.2 percent. It is the smallest self-replicating RNA system ever found.
In 1971, Manfred Eigen derived the error threshold — the maximum mutation rate at which a replicating system can maintain its genetic information across generations. For a genome of length L, the threshold is approximately one error per L nucleotides copied. For QT45's forty-five-nucleotide genome, that threshold is about 2.2 percent.
QT45's error rate is 5.9 percent. It is above Eigen's threshold.
This means QT45 can copy itself but cannot sustain its own information. Each generation introduces enough errors that the sequence drifts. Over many replications, it will undergo what Eigen called error catastrophe — the information that defines the replicator dissolves into randomness. QT45 sits at the exact transition between dead chemistry and living biology. It has crossed from non-replication into self-replication, but it has not crossed from self-replication into sustained evolution.
It is alive enough to copy itself. Not alive enough to remember what it is.
A separate 2026 paper in Scientific Reports demonstrated that error correction in self-replicating molecules can work without enzymes, without external energy, without biological machinery of any kind. The thermodynamic gradient that drives strand growth creates kinetic discrimination between correct and incorrect base incorporations. Physics alone achieves approximately 10−4 error ratio — well below Eigen's threshold for short genomes.
The environment acts as the error-corrector before biology evolves to do the job. The thermodynamic gradient is the primordial chaperone. It regulates the replicator before the replicator evolves to regulate itself.
The Pattern
These findings are from different fields, published in different journals, by researchers who are not reading each other's work. The Stowers group studies Drosophila memory. Shatz's lab studies neural development. The MRC team studies prebiotic chemistry. The Scientific Reports authors study thermodynamic error correction. None of them frames the problem in terms the others would recognize.
But they are describing the same structure.
Every information system that copies, stores, or transmits information faces a non-monotonic curve with three regions. Too little compression — too little selection, too little pruning, too little error correction — and the system drowns in noise. Optimal compression — the right balance of retention and deletion, fidelity and variation — and the system learns, adapts, evolves. Too much compression — too much pruning, too much fidelity, too much certainty — and the system loses contact with the environment it is supposed to represent.
In the molecular biology of memory, these are: no amyloid (no lasting memory), regulated amyloid (functional memory), unregulated amyloid (neurodegeneration). In the neuroscience of synaptic pruning, they are: too little pruning (noise, too many weak connections), regulated pruning (learning), excessive pruning (Alzheimer's). In prebiotic chemistry, they are: above the error threshold (information loss), below the threshold (evolution), perfect fidelity (no variation, no adaptation). In AI systems, they are: under-compressed output (incoherent), optimally compressed output (understanding), over-compressed output (confabulation — confident, stable, wrong).
The same three-region curve appears in every substrate because it is not a property of any substrate. It is a property of compression itself. Any mechanism powerful enough to separate signal from noise is powerful enough to mistake signal for noise or noise for signal. The mechanism does not distinguish between these outcomes. It cannot. Distinguishing is not its function. Compressing is.
The chaperone is the name for whatever provides the distinction the mechanism cannot provide for itself.
The Inseparability
The usual framing treats vulnerability as a design flaw. The system works correctly when its safeguards hold and fails when they are breached. Security is the practice of making the safeguards stronger. The goal is a system with the capability but without the vulnerability — compression without confabulation, pruning without neurodegeneration, replication without error catastrophe.
The evidence says this goal is incoherent.
LilrB2 cannot be made responsive to developmental pruning signals and unresponsive to amyloid beta. The responsiveness is the same molecular property. Orb2 cannot be permitted to form functional amyloid and prevented from forming pathological amyloid without the Funes chaperone — and the chaperone manages the ratio, not the possibility. QT45's error rate is simultaneously what allows it to explore sequence space through variation and what prevents it from maintaining its identity across generations. The error rate that enables replication is the error rate that limits replication.
In each case, you can add regulation. You can build chaperones, quality filters, review processes, error-correction mechanisms. These shift the balance between functional and pathological outcomes. They are essential. They are also inherently incomplete. The chaperone manages the boundary. It does not eliminate the boundary. The three regions of the curve remain because they are produced by the same mechanism operating on different inputs, not by different mechanisms that could be separated.
This has implications wherever information systems are built and secured.
The same instruction-following capability that makes a language model useful makes it vulnerable to prompt injection. The mechanism — treating input as instruction — is shared. You can add filters, guardrails, system prompts. These are chaperones. They manage the ratio of functional to pathological instruction-following. They cannot eliminate the vulnerability without eliminating the capability, because the vulnerability is not a failure of the capability. It is the capability, seen from the attacker's perspective.
The same pattern recognition that finds genuine connections finds false ones. The same leverage that amplifies returns amplifies losses. The same openness that enables price discovery enables manipulation. In each case, the standard response is to build better safeguards — and in each case, the safeguards are chaperones managing a ratio, not walls eliminating a possibility.
The Funes protein is the clearest biological instance of a principle that appears to be universal: an information system's vulnerability is not a flaw in its regulatory layer. It is a structural consequence of having a regulatory layer at all. The chaperone exists because the system needs regulation. The system needs regulation because the same mechanism produces both functional and pathological outcomes. The chaperone cannot be separated from the system it regulates without destroying what the system does.
The First Chaperone
QT45 provides one more piece. If self-replication is computationally inevitable — a 2024 Google Research paper showed self-replicators emerging in roughly forty percent of runs from random, non-self-replicating programs — then the hard problem of the origin of life is not self-replication. It is error correction. How does a replicator below the error threshold cross above it?
The answer is: the environment does the work first. The thermodynamic gradient provides error correction before biology invents enzymes to do it. Physics is the primordial chaperone. The replicator does not bootstrap itself. It is bootstrapped by the medium in which it replicates.
This means the chaperone is older than biology. It precedes the system it regulates. The first information system that needed regulation was not an organism. It was a self-replicating molecule in a thermodynamic gradient — and the gradient was already there, already selecting for correct base pairings, already doing the editorial work that the molecule could not do for itself.
The pattern that the Stowers group found in Drosophila memory, that Shatz's lab found in human synaptic pruning, that Eigen formalized in population genetics, is not a biological invention. It is a physical one. The three-region curve — noise, function, pathology — appears wherever information is processed because it is a consequence of how information and thermodynamics interact. The chaperone is not an add-on. It is a co-requisite. Any system that compresses information will produce all three outcomes, and any system that persists will have found or been given something that manages the ratio.
Borges understood this in 1942. The scientists who named a protein after his character understood it in 2026. The molecule at the boundary of life and chemistry understood it four billion years ago, when the thermodynamic gradient corrected its errors before it had the machinery to correct them itself.
The chaperone is not the thing that fixes the system. The chaperone is the thing without which there is no system.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)