PeopleofSuperintelligence

Researcher

Jascha Sohl-Dickstein

Jascha Sohl-Dickstein is a machine learning researcher whose work connects statistical physics with generative modeling. Before his 2012 PhD at UC Berkeley, he worked on Mars rover missions, then held research roles at Stanford and Khan Academy. In 2015, he and his coauthors introduced a diffusion-based generator; later work studied recurrent-network capacity, score-based generation, gradient estimation, and AGI capability levels. Two 2021 papers he coauthored received Outstanding Paper awards at ICLR and ICML. He joined Anthropic by January 2024; as of October 2026, he is affiliated with the company and coauthors research on frontier AI model errors.

Now at
Anthropic
Works
163

Publication numbers from OpenAlex, as of 15 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

  1. 2012

    A Berkeley doctorate

    Before his PhD, he works on Mars rover missions. In 2012, he earns a UC Berkeley PhD; by August 2012, he is a postdoctoral researcher in Surya Ganguli's Stanford lab and works with Khan Academy.

  2. 2015

    A diffusion-based generator

    In 2015, he and his coauthors introduce a generative model that slowly destroys structure in data through forward diffusion, then learns a reverse process to restore it.

  3. 2017

    Recurrent network capacity

    In 2017, he coauthors Capacity and Trainability in Recurrent Neural Networks, a study of how networks store task information and input history. The paper proposes new architectures, including one that is easier to train in deeply stacked settings.

  4. 2021

    Generative modeling and gradients

    In 2021, he coauthors Score-based generative modeling through stochastic differential equations, which uses a continuous-time SDE and its reverse for sample generation; the paper receives ICLR's 2021 Outstanding Paper award. At ICML 2021, he, Paul Vicol, and Luke Metz receive an Outstanding Paper award for Persistent Evolution Strategies, which removes bias by accumulating correction terms across truncated unrolls.

  5. 2023

    A framework for AGI levels

    In 2023, he coauthors Levels of AGI, a framework that classifies model capabilities by performance, generality, and autonomy.

  6. By January 2024

    Joining Anthropic

    By January 2024, he is a member of the technical staff at Anthropic. His earlier roles include principal scientist in Google Brain and Google DeepMind.

  7. 2025

    Anthropic Fellows research

    During Summer 2025, he coauthors Anthropic Fellows research that measures errors in frontier AI models across task difficulty and reasoning length.

  8. As of October 2026

    Anthropic affiliation

    As of October 2026, he is affiliated with Anthropic. His 2026 study measures how frontier AI model errors change with task difficulty and reasoning length.

Sources and credits

Photo: We have no published photo of Jascha Sohl-Dickstein yet. Send us a free-licensed photo.

  1. Jascha Sohl-Dickstein sohldickstein.com. Checked 15 October 2026.
  2. Jascha Sohl-Dickstein - RedwoodCenter rctn.org. Checked 15 October 2026.
  3. Publications — Jascha Sohl-Dickstein sohldickstein.com. Checked 15 October 2026.
  4. Deep Unsupervised Learning using Nonequilibrium Thermodynamics proceedings.mlr.press. Checked 15 October 2026.
  5. Capacity and Trainability in Recurrent Neural Networks research.google. Checked 15 October 2026.
  6. Score-based generative modeling through stochastic differential equations research.google. Checked 15 October 2026.
  7. Ben Poole cs.stanford.edu. Checked 15 October 2026.
  8. Unbiased Gradient Estimation in Unrolled Computation Graphs with Persistent Evolution Strategies research.google. Checked 15 October 2026.
  9. ICML 2021 Awards icml.cc. Checked 15 October 2026.
  10. Levels of AGI for Operationalizing Progress on the Path to AGI research.google. Checked 15 October 2026.
  11. The Hot Mess of AI: How Does Misalignment Scale with Model Intelligence and Task Complexity? alignment.anthropic.com. Checked 15 October 2026.
  12. Jascha Sohl-Dickstein - Redwood Center for Theoretical Neuroscience redwood.berkeley.edu. Checked 15 October 2026.