Daniel Levy
Last reviewed
Jun 5, 2026
Sources
21 citations
Review status
Source-backed
Revision
v2 · 2,348 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 5, 2026
Sources
21 citations
Review status
Source-backed
Revision
v2 · 2,348 words
Add missing citations, update stale details, or suggest a clearer explanation.
Daniel Levy is an artificial intelligence researcher and co-founder and president of Safe Superintelligence Inc. (SSI), the AI laboratory he started in June 2024 with former OpenAI chief scientist Ilya Sutskever and the investor and technologist Daniel Gross [1][2][3]. Before co-founding SSI, Levy led the optimization team at OpenAI, where he was also a core technical lead on GPT-4, and he completed a PhD in computer science at Stanford University, where his research centered on optimization, differential privacy, and the methods used to train large-scale machine learning models [4][5][6][16].
Levy works at the intersection of machine learning theory and large-scale model training. His academic output focused on stochastic optimization, distributionally robust optimization, and privacy-preserving learning, and during his time in industry he applied that background to the engineering of large neural networks [4][5]. He is most prominent today for his role at Safe Superintelligence, where he is one of three co-founders and, since July 2025, holds the title of president [2][3]. Press coverage of SSI's founding has described Levy as an AI researcher, distinguishing his technical profile from that of his co-founder Daniel Gross, who is better known as an investor and former head of Apple's AI efforts [1][7].
Because "Daniel Levy" is a common name, the figure described here is specifically the SSI co-founder whose Stanford and OpenAI affiliation is documented on his Stanford AI Lab page and in his published research record [4][8]. His full name in academic publications is given as Daniel Asher Nathan Levy [8]. He should not be confused with the English businessman Daniel Levy, the long-serving executive chairman of the football club Tottenham Hotspur and managing director of the investment company ENIC, who is a different and unrelated person [17].
Levy studied at Stanford University, where he earned a master's degree and then a PhD in computer science [4][8]. He was a master's student from 2015 to 2018, advised by Stefano Ermon, working on probabilistic models and reinforcement learning, and a PhD student from 2018 to 2022, advised by John Duchi [4][8]. His own curriculum vitae dates the master's program to September 2015 through June 2018 (with a reported grade point average of 4.04) and the doctoral program to September 2018 through December 2021 [6]. His doctoral dissertation, "Advancing optimization for modern machine learning," was completed in 2021 and addressed adaptive gradient algorithms, robustness to out-of-distribution inputs, and private optimization for deployed models [9][16].
Before Stanford, Levy studied in France. His Stanford page lists a Bachelor of Science and Master of Science from the École Polytechnique, obtained in 2014 and 2015, and earlier study at the Lycée Louis-le-Grand in Paris [4]. His curriculum vitae records the École Polytechnique credential as a Diplome d'ingenieur earned between September 2012 and July 2015, and notes that he ranked 13th nationally on the school's competitive entrance examination [6]. Polytechnique is one of France's most selective institutions for science and engineering, and the two preceding years at the Lycee Louis-le-Grand (2010 to 2012) were spent in the intensive preparatory program, on a mathematics, physics, and computer science track, that French students complete before sitting the grandes ecoles entrance exams [6].
During his studies Levy completed a series of research internships in both France and the United States. His curriculum vitae lists early internships at Microsoft in Paris in 2014, where he worked on analytics for the Cosmos big-data platform and a churn-prediction project for the Xbox Music service, and at the machine-learning startup Shift Technology in Paris in 2015, where he developed bandit methods for fraud detection that the company put into production for insurers [6]. He then interned at Facebook's Applied Machine Learning group in Menlo Park in 2016 (bandit and reinforcement-learning methods for active learning), at Google Brain in Mountain View in 2017 (Markov chain Monte Carlo methods, with Jascha Sohl-Dickstein and Matt Hoffman), and at Google Research in New York in 2020 (differential privacy, with Ananda Theertha Suresh, Satyen Kale, and Mehryar Mohri) [4][6]. He was also a teaching assistant at Stanford for the convex optimization course EE364A and the machine learning course CS229 [6].
| Stage | Institution | Period | Advisor or focus |
|---|---|---|---|
| Preparatory studies | Lycée Louis-le-Grand, Paris | 2010-2012 | mathematics and physics |
| B.S. and M.S. / Diplome d'ingenieur | École Polytechnique | 2012-2015 | engineering and applied science; ranked 13th nationally |
| M.S. (computer science) | Stanford University | 2015-2018 | Stefano Ermon; probabilistic models, reinforcement learning |
| PhD (computer science) | Stanford University | 2018-2022 | John Duchi; optimization, privacy |
Levy's research spans optimization, robustness, and privacy in machine learning. His doctoral work studied how to make optimization methods adapt automatically to the difficulty of a problem, reducing the need for manual hyperparameter tuning, and how to train models that remain reliable when the data they see at deployment differs from the data they were trained on [9]. A representative theoretical paper, "Necessary and Sufficient Geometries for Gradient Methods," was selected for an oral presentation at NeurIPS 2019 [4]. According to Levy's curriculum vitae, that talk was one of 36 oral presentations chosen from 6,743 submissions, placing it in roughly the top half of one percent of accepted work [6]. The paper characterizes the problem geometries for which standard and adaptive (diagonally preconditioned) stochastic gradient methods are minimax optimal, helping to explain when adaptive methods such as AdaGrad-style algorithms are theoretically justified [18].
A second strand of his work concerns distributionally robust optimization, which seeks models that perform well across a range of possible data distributions rather than only the training distribution. His 2020 paper "Large-Scale Methods for Distributionally Robust Optimization," written with Yair Carmon, John Duchi, and Aaron Sidford, gave practical stochastic algorithms that scale this idea to large datasets, and the 2021 paper "Distributionally Robust Multilingual Machine Translation" applied the same principles to balancing performance across many languages in a translation model [8][6]. He also published on differential privacy, including "Learning with User-Level Privacy" and "Adapting to function difficulty and growth conditions in private optimization," both appearing at NeurIPS in 2021 [8][6]. The user-level privacy work studies how to learn accurately when the privacy guarantee must protect everything contributed by a single user rather than a single record, a setting relevant to training on data pooled from many individuals [6].
Earlier in his career Levy contributed to several widely cited papers in deep learning. "Generalizing Hamiltonian Monte Carlo with Neural Networks" (ICLR 2018), written with Matthew Hoffman and Jascha Sohl-Dickstein, introduced a way to use neural networks to learn better Markov chain Monte Carlo samplers and has accumulated well over one hundred citations [19]. As a coauthor on "Data Noising as Smoothing in Neural Network Language Models" (ICLR 2017), a project that also involved Andrew Ng and Dan Jurafsky, he worked on regularization techniques for language models [6]. His earlier work also touched on encrypted inference, with "Faster CryptoNets: Leveraging Sparsity for Real-World Encrypted Inference," and on reinforcement learning for discrete action spaces with "Deterministic Policy Optimization by Combining Pathwise and Score Function Estimators" [8][6].
| Year | Work | Topic |
|---|---|---|
| 2017 | Data Noising as Smoothing in Neural Network Language Models (ICLR) | language model regularization |
| 2018 | Generalizing Hamiltonian Monte Carlo with Neural Networks (ICLR) | learned MCMC sampling |
| 2018 | Faster CryptoNets: Leveraging Sparsity for Real-World Encrypted Inference | privacy-preserving inference |
| 2019 | Necessary and Sufficient Geometries for Gradient Methods (NeurIPS oral) | optimization theory |
| 2020 | Large-Scale Methods for Distributionally Robust Optimization (NeurIPS) | robust optimization |
| 2021 | Learning with User-Level Privacy (NeurIPS) | differential privacy |
| 2021 | Distributionally Robust Multilingual Machine Translation (EMNLP) | robust optimization, NLP |
| 2021 | Advancing optimization for modern machine learning (PhD thesis) | optimization for ML |
After completing his PhD, Levy joined OpenAI, where he led the company's optimization team [4]. His curriculum vitae records that he joined OpenAI in San Francisco as a Member of Technical Staff in March 2022, "Leading the Optimization team," and his Stanford AI Lab page likewise states, "I currently lead the Optimization team at OpenAI" [4][6]. Optimization in this setting covers the algorithms and engineering used to train very large neural networks efficiently and stably, an area closely connected to his graduate research on gradient methods and large-scale learning [4][9].
Levy was a core technical contributor to GPT-4, OpenAI's flagship large language model released in March 2023. In the contributions section of the GPT-4 Technical Report he is credited as one of two "overall vision co-lead[s]" and as the project's "optimization lead," and he is also listed among the contributors to "training run babysitting" and to "deployment and post-training" [18]. He subsequently served as the optimization lead for GPT-4V, the vision-enabled version of the model [20]. These roles placed him among the small group responsible for the stability and efficiency of one of the largest training runs of its era. He left OpenAI in 2024 to help found Safe Superintelligence [1][2].
On June 19, 2024, Ilya Sutskever announced Safe Superintelligence Inc., naming Daniel Levy and Daniel Gross as his co-founders [1][2]. Sutskever, a co-founder and former chief scientist of OpenAI, had left that company earlier in 2024; Gross had previously led artificial intelligence efforts at Apple and was a well-known startup investor [1][7]. The new venture set up offices in Palo Alto, California, and Tel Aviv, Israel [1][2].
At founding, Levy was one of the three co-founders. The company's leadership structure changed in mid-2025. On about June 29, 2025, Daniel Gross departed SSI, with reporting indicating he moved to Meta's superintelligence effort, and Sutskever took over as chief executive officer [2][3][10]. Gross joined what became Meta Superintelligence Labs, the unit Mark Zuckerberg assembled in 2025 [21]. At the same time, according to Sutskever, Levy became president of the company [3][10]. TechCrunch reported that "Safe Superintelligence co-founder Daniel Levy is becoming president of the startup, according to Sutskever," while Sutskever continued to oversee the technical team in addition to his new chief-executive duties [3]. Reporting from early 2026 continued to describe SSI as operating "under the leadership of Sutskever and President Daniel Levy" [21].
Safe Superintelligence describes itself as "the world's first straight-shot SSI lab," pursuing what it calls "one goal and one product: a safe superintelligence" [11]. Sutskever has said the company's "first product will be the safe superintelligence, and it will not do anything else up until then," signaling that, unlike OpenAI, SSI does not plan to ship intermediate commercial products before reaching its objective [1]. The company frames building safe superintelligence as the most important technical problem of its time [11].
Despite having no released product, SSI has raised large sums and reached a high valuation. In September 2024 it announced about $1 billion in funding at a reported valuation near $5 billion, with backers including Andreessen Horowitz, Sequoia Capital, DST Global, SV Angel, and the investment partnership NFDG run by Nat Friedman and Daniel Gross [1][12][13]. In a round reported in March and April 2025, SSI raised roughly $2 billion at a valuation of about $30 billion to $32 billion, led by Greenoaks Capital [14][15]. In April 2025, Google Cloud announced a partnership to supply the company with tensor processing units for its research [2]. By early 2026 the company was still reported to carry a valuation of about $32 billion and to have raised more than $3 billion in total, having reportedly turned down an acquisition offer from Meta in June 2025 [21].
| Date | Funding event | Reported valuation | Selected investors |
|---|---|---|---|
| June 19, 2024 | Company founded | n/a | founders Sutskever, Gross, Levy |
| September 2024 | ~$1 billion raised | ~$5 billion | a16z, Sequoia, DST Global, SV Angel, NFDG |
| March-April 2025 | ~$2 billion raised | ~$30 billion to $32 billion | Greenoaks Capital |
Levy's curriculum vitae lists several distinctions from his academic career [6]. He was selected for the Google Brain Residency Program in 2017, a position his CV describes as drawing roughly the top one percent of applicants. His NeurIPS 2019 paper was chosen for an oral presentation, an honor extended to about the top half of one percent of submissions that year. He was a finalist for a Facebook Fellowship in 2020, and he was nominated by Stanford University for a Google PhD Fellowship, a nomination the university limited to two students [6]. He was also ranked 13th nationally on the École Polytechnique entrance examination [6].