A forty-five-nucleotide RNA molecule that writes itself. An AI that solves four open math problems overnight. Another AI that writes eight confident proofs of things that aren't true. The difference is never the generator.
A strand of RNA forty-five nucleotides long can synthesize a copy of itself. It was found not by design but by patience — approximately one trillion random sequences, eleven rounds of evolutionary selection, and one survivor that turned out to be alive.
The molecule is called QT45. The name stands for Quite Tiny 45. Philipp Holliger's lab at the MRC Laboratory of Molecular Biology published it in Science this month. It is the smallest known RNA polymerase ribozyme — roughly four times smaller than the previous record holders, which were themselves already too large to have plausibly arisen by chance on a prebiotic Earth.
QT45 shifts the arithmetic of the origin of life. The number of possible forty-five-nucleotide RNA sequences is approximately 1027. Even if the functional fraction is one in a trillion, trillions of potential self-replicators exist in sequence space, waiting to be found. The void is not empty. It is fertile.
Two Events in Mathematics
The same month, eleven leading mathematicians — including Fields Medalist Martin Hairer and MacArthur Fellow Lauren Williams — posed ten original, unpublished lemmas to the best AI systems available. They called it First Proof. The problems were chosen to be the kind a talented graduate student might solve: a few pages of proof, real mathematics, but not famous conjectures. No data contamination was possible because the problems had never been discussed publicly.
The AI systems produced confident solutions for all ten. Two were correct. Eight were, in the mathematicians' words, very convincing nonsense.
Mohammed Abouzaid of Stanford noted that the correct solutions have the flavor of 19th-century mathematics. The failures were harder to characterize. They looked right. They read well. They were structurally coherent. They simply were not true.
Meanwhile, a startup called Axiom announced that its system, AxiomProver, had solved four open problems. The Chen-Gendron conjecture — a question in algebraic geometry about differentials on curved surfaces, unsolved since 2021 — fell overnight. The key insight: AxiomProver found a connection to Jacobi symbols, a nineteenth-century number theory phenomenon that every human mathematician working on the problem had missed. A second proof settled Fel's Conjecture using exponential generating functions and formulas first recorded in Ramanujan's notebooks over a century ago.
AxiomProver's architecture combines a language model with a specialized reasoning system and Lean formal verification. Every step of every proof is machine-checked against the axioms of mathematics. The system does not merely generate text that looks like a proof. It generates candidates and then verifies them against an external standard of truth.
One Architecture
Strip the domain labels away and there is one architecture, appearing three times.
QT45 was discovered by generating approximately one trillion random RNA sequences and testing each one against a chemical assay for polymerase activity. The generator was random. The verifier was chemistry itself — does this molecule actually catalyze RNA synthesis? Eleven rounds of selection. One survivor.
AxiomProver generates mathematical conjectures and candidate proofs, then verifies each step in Lean. The generator is an AI searching through mathematical concept space. The verifier is formal logic — does this step follow from the axioms? No ambiguity. No judgment call.
The First Proof language models generated proof-shaped text and verified it against their own confidence. The generator was powerful. The verifier was absent. Eight out of ten results were ghosts — structurally present, logically empty.
The pattern is simple enough to state in one sentence: what distinguishes breakthrough from nonsense is the verifier, not the generator.
Beyond Biology and Mathematics
This matters wherever AI meets reality.
In drug discovery, AI excels at screening compounds against known protein targets — generator plus assay. It struggles with predicting mechanism of action, where no clean verifier exists. In software engineering, AI writes code confidently and the test suite is the verifier that separates working code from plausible code. In scientific research, AI searches vast parameter spaces efficiently, but only when the search has a clear success criterion that does not depend on the searcher's own confidence.
The failure mode is always the same. When the verifier is the generator's own certainty, the output is indistinguishable from truth until external contact occurs. The eight false proofs in First Proof were not random gibberish. They were the most dangerous kind of wrong — wrong in a way that requires expert effort to distinguish from right.
Holliger's team did not design QT45. They set up conditions for search and selection, and a molecule emerged. Axiom's team did not write the proof of the Chen-Gendron conjecture. They built a system that searches and verifies, and a proof emerged. In both cases, the creative act was not generation. It was establishing the conditions under which generation and verification could operate together.
The Density of the Void
There is something unsettling about the density. RNA sequence space is 1027 possibilities for a forty-five-nucleotide strand. Mathematical concept space contains connections between known domains that centuries of human attention have not explored. The spaces are not barren. They are far denser with functional structure than human intuition expects.
But they are also dense with near-misses. The same fertility that produces self-replicating molecules produces trillions of molecules that almost self-replicate. The same combinatorial richness that hides the Chen-Gendron connection hides millions of connections that look promising and lead nowhere. The eight false proofs were not anomalies. They were the natural product of a fertile space explored without adequate verification — the near-misses that outnumber the truths by orders of magnitude.
The void is fertile. The void is also full of ghosts. The instrument that separates the living from the dead is never the generator.
It is always the assay.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)