'Thermodynamic computer' can mimic AI neural networks — using orders of magnitude less energy to generate images

Researchers have created a "generative thermodynamic computer" capable of producing images from random disturbances (thermal noise), effectively mimicking the generative AI functions of neural networks but using orders of magnitude less energy. Unlike conventional computing, which minimizes thermal noise, this new system harnesses it, similar to how a surfer uses wave power. The work, detailed in *Physical Review Letters*, draws inspiration from diffusion models used in AI, where images are reconstructed from noise. By manipulating the Langevin equation, the scientists calculated the necessary coupling strengths (like circuit connection strengths) to reverse the noise process step-by-step, generating an image from random input. This probabilistic computing approach offers significant efficiency gains, particularly for optimization problems, and provides a fundamental physical interpretation to a field often dominated by "black-box" models, though the current demonstration is rudimentary.

(Source：Live Science)