Research

Bridging AI reasoning with
the physical world

Normal Computing was founded by former members of Google Brain and Google X who helped first pioneer reliable deep learning to mission-critical production systems, and developed the leading ML frameworks for Probabilistic and Quantum AI. Normal began with the goal of solving the energy crisis of intelligence by developing unconventional chip architectures which accelerate probabilistic AI reasoning algorithms.

We believe that exploring the limits of fundamentally new chip ideas, including those which optimize their own physics, requires the help of AI, and ultimately the realization of a virtuous cycle first detailed by Von Neumann and best depicted by the Ouroborus: AI reasoning for hardware and hardware for AI reasoning.

Sound On

Algorithms and AI Engineering

Our software and algorithms research focuses on advancing AI reasoning from two compatible perspectives:

First, we are developing cutting-edge auto-formalizing AI systems for silicon inspired by proof-solving models like Google DeepMind's AlphaGeometry, leveraging advanced reinforcement learning techniques, and curated and synthetic data, to combine LLMs with Formal Models for complex reasoning and deductive logic.

Second, we are pioneering a probabilistic perspective to AI decision-making. The real world – like any data – is noisy and incomplete, and so fundamentally probabilistic. And thus navigating it efficiently requires a principled view of uncertainties so that one can reliably reason and explain across complex information and insight pathways without hallucination. This entails a fundamental upgrade to existing AI systems – probabilistic thinking – to reliably augment human intelligence at scale.

Thermodynamic Computing and Silicon

We are scaling down AI costs 1000x with new chips for some of the hardest and largest-scale reasoning problems in AI.

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.

Thermodynamic computing was pioneered at Normal by folks including Dr. Patrick Coles and Dr. Gavin Crooks. We publish in top-tier venues and are proudly supported by the public sector.

READ THE latest research

Our research team is behind the innovations that enabled scaling probabilistic programming and probabilistic machine learning (TensorFlow Probability, Stan), and the modern approach to near-term quantum computation (Tensorflow Quantum, NISQ). Together, the Normal team has pioneered Thermodynamic AI, a physics-based computing paradigm for accelerating the key primitives in probabilistic machine learning.

We also created Posteriors and Outlines, groundbreaking open-source frameworks for uncertainty quantification and controlled LLM generations, respectively, and maintain DSPy, the leading framework for language model program optimization.

5.19.2025

Prototype Computer Uses Noise to Its Advantage

5.19.2025

Prototype Computer Uses Noise to Its Advantage

3.27.2025

ChatGPT's Hunger for Energy Could Trigger a GPU Revolution

With AI projects booming and the physical limits of silicon looming, some startups are challenging Nvidia's dominance and say it’s time to reinvent the computer chip entirely.

12.23.2024

How to fix computing's AI energy problem: run everything backwards

10.31.2024

Normal Computing Selected for ARIA's £50M Scaling Compute Programme to Revolutionize AI Hardware Costs

5.22.2024

Introducing Fuji-Web

Fuji-Web is an intelligent AI partner that understands the user’s intent, navigates websites autonomously, and executes tasks on the user’s behalf while explaining each action step.

4.16.2024

posteriors: Normal Computing’s library for Uncertainty-Aware LLMs

Online learning and hallucination detection are understood as frontier problems in AI and with LLMs. posteriors is a new open source Python library from Normal Computing that provides tools for uncertainty quantification and Bayesian computation using PyTorch and its functional API.

1.23.2024

Normal Computing Unveils the First-ever Thermodynamic Computer

1.16.2024

First 'thermodynamic computer' uses random noise to calculate

11.9.2023

A First Demonstration of Thermodynamic Matrix Inversion

Thermodynamic computing offers a natural approach for fast, energy-efficient computations. We report on the first-ever experiment towards thermodynamic artificial intelligence: solving matrix inversion problems by allowing a system of coupled electrical oscillators to thermally equilibrate with its environment.

11.5.2023

Developing Advanced Reasoning and Planning Algorithms with LLMs

In this post we introduce Branches, our tool for prototyping and visualizing advanced LLM reasoning and planning algorithms. We apply Branches to the problem of generating Python code for HumanEval.

10.24.2023

Supersizing Transformers: Beyond RAG with Extended minds for LLMs

In this blog we discuss how the transformer architecture naturally extends over external memories, and share empirical results which leverage this capability. These methods are innate (don't require fine tuning) and outperform popular retrieval augmented generation methods.

10.20.2023

Explainable Language Models: Existing and Novel Approaches

We review key aspects of explainability for language models and introduce some Normal innovations.

9.11.2023

Computers that use heat instead of electricity could run efficient AI

8.4.2023

Eliminating hallucinations (fast!) in Large Language Models with Finite State Machines

In this blog, we introduce our method for regex-guided generation implemented in Outlines

6.13.2023

Normal Computing Raises $8.5M in Seed Funding to Enable AI Solutions For Critical Enterprise and Government Applications

Join us in redefining  the future of computing.