Large language models changed how people interact with information. They can summarize, explain, search, code and reason through text at extraordinary scale. But some of the world’s most valuable decisions are not language problems. They are governed by quantities.
How much capital is at risk? Which molecule should be synthesized? Which alloy should be tested? What operating point maximizes throughput without violating a constraint? Which experiment will reduce uncertainty the most?
For these questions, plausibility is not enough. A quantitative system has to live in a world of units, boundary conditions, distributions, equations, constraints, measurement error and physical consequences.
The public market is already moving in this direction. SandboxAQ describes Large Quantitative Models for scientific and materials problems. PhysicsX builds Large Physics Models from high-fidelity simulation data. Siemens combines AI with comprehensive digital twins. Schrödinger combines physics-based molecular modelling, machine learning and scientific workflow software.
The common direction matters more than any one company’s terminology: AI is moving from generating descriptions of the world to helping model the world itself.
The infrastructure requirement changes
A useful quantitative AI system needs more than inference. It needs data lineage, model identity, simulation and solver interfaces, uncertainty, calibration, constraints, result ingestion, reproducibility, evidence and benchmarking.
If the result controls capital, an experiment, a physical process or a scientific claim, the system should answer not only “what is the number?” but also “why should anyone trust this number?”
That is the infrastructure layer LargeQuant is building.