Scientists have demonstrated a “thermodynamic computer” that performs tasks similar to AI neural networks but uses orders of magnitude less energy by harnessing natural thermal fluctuations instead of fighting them, tells Live Science. Traditional AI systems require vast power to drive digital chips and suppress noise. This new approach lets the unavoidable jiggle of atoms and molecules above absolute zero do much of the work. By treating this noise as a computational resource, researchers showed they could generate images in a way that mirrors diffusion-based AI models while consuming far less energy.
The concept builds on probabilistic computing principles where systems settle into equilibrium states that encode solutions. In practice, the team at Lawrence Berkeley National Laboratory used a thermodynamic model to reverse the effect of noise on simple images of handwritten numerals, similar to how modern generative AI models remove noise to create structured outputs. The simulations leveraged the Langevin equation, a century-old physics model, to describe and reverse how noise alters data.
This method stands in contrast to mainstream AI, which expends energy to sharply define every bit of data and suppress random physical effects. By contrast, thermodynamic computing embraces randomness, yielding a natural path to solve optimization problems with much lower energy cost. Researchers outside the immediate project have built low-energy circuits that exploit similar principles for basic computation tasks.
The work remains in early stages. Image generation demonstrated so far is rudimentary, and researchers acknowledge significant challenges in scaling thermodynamic hardware to match modern AI models’ performance. Still, this line of research points to a future where physics-inspired computing offers a deeply interpretable alternative to power-hungry “black box” AI systems and could shift how machine learning tasks are accomplished at the hardware level.