Daniel Levy

Daniel Levy trained in France, at the Lycée Louis-le-Grand preparatory program and then École Polytechnique, where the degree was in applied mathematics and computer science, before moving to Stanford in 2015. A master's with Stefano Ermon covered probabilistic models and reinforcement learning, and an early paper on smoothing neural language models by adding noise to their training data (ICLR 2017) had Andrew Ng among its co-authors. A 2017 internship at Google Brain with Jascha Sohl-Dickstein and Matt Hoffman produced "Generalizing Hamiltonian Monte Carlo with Neural Networks" (ICLR 2018), which trained a neural network to make a standard sampling algorithm mix faster. Levy's PhD with John Duchi, finished in December 2021, was on optimization for machine learning, including training methods that minimize loss on the worst-served slices of a dataset and differentially private learning.

Levy joined OpenAI in March 2022 and led its optimization team, the group that tunes how very large models are trained. The GPT-4 technical report credits Levy as optimization lead and as a co-lead of the overall vision work that let GPT-4 accept images, and also lists Levy among the staff who watched over the main training run.

On 19 June 2024 Levy, Ilya Sutskever and Daniel Gross announced Safe Superintelligence Inc., a lab with "one goal and one product." When Gross left for Meta at the end of June 2025, Sutskever became CEO and told staff and investors that "Daniel Levy is President." Nvidia's July 2026 announcement of its partnership with SSI described the company as "led by Ilya Sutskever and Daniel Levy."