A foundational pillar of artificial intelligence alongside linguistic intelligence.
“There are, indeed, things that cannot be put into words. They make themselves manifest.”
— Ludwig Wittgenstein, Tractatus Logico-Philosophicus, 6.522
What it means
Intelligence is the ability to acquire knowledge, adapt to unfamiliar situations, and solve new problems. While modern AI has largely focused on language, true intelligence is not confined to words. Numerical intelligence extends this cognitive capacity directly into the measurable world. It is the ability to perceive raw numerical data, instantly deduce the underlying mechanisms, and reason about complex systems without relying on linguistic translation.
Traditional deep learning often absorbs a specific system's rules directly into the model’s weights. This approach effectively memorizes a fixed scenario and demands costly retraining for any new application. Numerical intelligence fundamentally breaks this mold. Like Large Language Models, it extracts knowledge instantly from numerical prompts. Its weights serve as a generalized reasoning engine, empowering the model to understand and predict entirely new systems based solely on the measurements supplied at inference, eliminating the need for retraining.
The systems that shape science, society, and business, from climate and energy to cities, markets, and supply chains, are continuously revealed through measurements. The ability to rapidly understand how these systems behave, predict what comes next, and guide action can turn changing data into timely decisions. Numerical intelligence is therefore indispensable to a future AGI ecosystem, giving artificial intelligence a direct capacity to understand, predict, and act on real-world systems.
Research foundation
Our proposed In-Context Operator Network (ICON) is an early demonstration of numerical intelligence for forecasting across scientific and social systems. Trained on datasets across diverse domains, including traffic, hydrology, power, and urban mobility, ICON can generalize to new disciplines entirely absent from its training data. By reading new data as a numerical prompt rather than requiring retraining, it adapts to unseen fields like web activity and air quality, and can even beat state-of-the-art models trained specifically for those tasks.
A dedicated inference-time program, acting as a harness similar to those in linguistic intelligence, orchestrates calls to a frozen numerical intelligence model alongside explicit mathematical operations. It transforms numerical prompts, quantifies uncertainty, and enforces hard rules. This programmable layer unlocks the full potential of the foundation model without a single weight update.
To accelerate the development of numerical intelligence, we built an automated research system where co-evolving agents iteratively improve core model architectures, refine training methods, and optimize the inference harness. This continuous self-improving loop dramatically scales our research capacity, allowing us to push the boundaries of numerical intelligence at AI speed.
