AI models differ from other computing by producing probabilistic output. A convolutional neural network (CNN) that classifies objects doesn’t definitively identify them; it only reports what it “thinks” is the most likely label given its training. User interfaces typically suppress the probability estimate, but it’s there. Other models use probabilistic inputs during training. Generative AI often employs a diffusion process that uses Gaussian noise.
Digital computers are deterministic, a desirable characteristic for classic tasks but unnecessary for AI models with probabilistic outputs and inputs. Quantum computing promises to be faster, but the technology has been slow to mature. Founded by quantum-computing refugees, startup Extropic is taking a different approach, developing processors that employ thermodynamic effects to speed up probabilistic machine learning.
Adoption of radical new technologies typically takes years if they prove viable—which itself is uncommon. Moreover, as hot a topic as generative AI is, it has yet to achieve mainstream adoption. Nonetheless, radical change like that Extropic proposes could lower the cost of generative AI, helping to make it a common tool.