Formal Logic Meets LLMs

We are building auto-formalizing systems that combine the deductive rigor of symbolic proofs with the flexible abstraction of large language models. Inspired by AlphaGeometry and similar systems, we use synthetic data and advanced reinforcement learning to align LLMs with Formal Models for complex reasoning and deductive logic in chip design from specifications.

Probabilistic Reasoning

The real world is uncertain, noisy, and bound to the laws of physics. To simulate and reason within it, AI systems must be fundamentally probabilistic. We're co-designing generative diffusion-like models, grounded in unstructured information, to reason coherently about the physical world, without hallucination.

Physics-Based Approach

Because probabilistic reasoning is ubiquitous in nature, we turn to the natural world for clues on how to build these kinds of AI computing systems in silicon using physics-based principles of thermodynamics and Bayesian learning.

Scaling Diffusion and SDEs

To reduce AI costs by orders of magnitude, we are developing new silicon built for scaling diffusion-like models. Our thermodynamic computing chips leverage principles like Langevin dynamics, non-equilibrium driving, and asynchronization to perform efficient stochastic differential equation emulation in physics.

The New "Fluctuation" Frontier

Thermodynamic computing was pioneered at Normal by the Founders alongside Dr. Patrick Coles and Dr. Gavin Crooks – recognized experts in quantum physics and statistical mechanics. We are supported by public sector programs and publish regularly in venues spanning AI, semiconductors, physics, and hardware systems.