creators

Jascha Sohl-Dickstein: diffusion models, Anthropic

Jascha Sohl-Dickstein is a member of technical staff at Anthropic and first author of the 2015 paper that introduced diffusion probabilistic models.

Jascha Sohl-Dickstein is an American machine learning researcher and a member of technical staff at Anthropic, which he joined in February 2024. He took his PhD in biophysics at UC Berkeley in May 2012, in Bruno Olshausen's lab at the Redwood Center for Theoretical Neuroscience, after two bachelor's degrees at Cornell and four years as a research associate on NASA JPL's Mars Exploration Rovers. He is the first author of "Deep Unsupervised Learning using Nonequilibrium Thermodynamics", posted in March 2015, which introduced diffusion probabilistic models by analogy with non-equilibrium statistical physics; the model family later became the basis of image, video and audio generation systems. He spent nine years at Google Brain and Google DeepMind, from 2015 to 2024, ending as a principal scientist, and worked there on the theory of infinite-width networks, on learned optimizers, and on BIG-bench, the large language model benchmark he co-led. At Anthropic he co-authored "The Hot Mess of AI" in February 2026, on how model failures shift from systematic misalignment towards incoherence as tasks grow harder.

Specialization

diffusion models, neural network theory, learned optimizers, AI alignment

Country

United States