Quantaeon

The Universal Quantum Transformer now speaks. We previously showed the UQT could learn exact mathematics (modular arithmetic and non-Abelian algebra) on a 5-qubit circuit with fewer than 700 parameters. Today, we present its results on language, and they are just as striking. The benchmark: SCAN "add jump." Train the model on "walk left," "run right," and "look around," but never show "jump" combined with any direction. Then ask it to classify "jump left" in zero-shot. Standard Transformers fail this. Despite hundreds of thousands of parameters, they oscillate between 0% and 100% at convergence. They never crystallize. The UQT, with 1,650 parameters on a 6-qubit register, crystallized into stable 100% zero-shot accuracy. Then we ran it on physical IBM Quantum hardware: 10/10 evaluations passed. Why does it work? Verbs and directions occupy orthogonal subspaces of the quantum state. The circuit does not interpolate. It physically superposes two independent geometric representations and reads both simultaneously. The physics composes the language. By construction. We have now validated crystallization across four domains: ⚛️ Modular addition — 551 params ⚛️ Modular multiplication — 551 params ⚛️ Non-Abelian permutations (S₄) — 690 params ⚛️ Linguistic compositionality (SCAN) — 1,650 params 40 hardware evaluations. 39/40 passed. 100% on every unseen generalization test. Classical networks approximate algebra with statistics. The UQT is algebra. Full blog post below. 👇 Blog post: medium.com/@quantaeon.ai/the…

For years, AI has moved in one direction: Larger models. Larger data centers. Greater dependence on the cloud. What if more intelligence does not require more scale? At @Quantaeon, we are building quantum-inspired AI for private, local, enterprise intelligence. Read more about our core technology: medium.com/@quantaeon.ai/why… @artalebpour @sungyongchung @IBMResearch @IBM