Ilya

Ilya Sutskever coauthored an end-to-end method for sequence-to-sequence learning, became OpenAI’s founding research director, and cofounded Safe Superintelligence Inc. (SSI). His recent work centers on investigating better generalization and building safe superintelligence; a July 2026 partnership with NVIDIA marked his stated decision that SSI’s research was ready to scale.[3][4][9][8]

SSI was founded in 2024. In a July 3, 2025 message, Sutskever announced that he was formally CEO, Daniel Levy was President, and the technical team continued to report to him. The message followed Daniel Gross’s departure, effective June 29. In his November 2025 interview with Dwarkesh Patel, Sutskever said Meta had offered to acquire SSI and he had declined.[8][7][9]

SSI’s stated mission is a single goal and product: safe superintelligence. It proposes advancing safety and capabilities together while insulating research from short-term commercial pressures. Sutskever described the company more tentatively in November 2025: an attempt to investigate promising ideas about generalization and discover whether they were correct. “We are squarely an ‘age of research’ company,” he said.[6][9]

In the July 27, 2026 partnership announcement, Sutskever said, “We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so.” NVIDIA announced an investment and access to its Vera Rubin platform intended to enable a tenfold increase in SSI’s compute. Sutskever called the platform choice a “big bet”; the expansion was prospective.[8]

In November 2025, Sutskever argued that pretraining had given researchers a reliable recipe: combine data, compute, and a neural network, then improve results by increasing scale. With finite data and already-large computers, he saw renewed importance in finding more productive training methods: “So it’s back to the age of research again, just with big computers.” His argument left room for further scaling while questioning whether enlarging the existing recipe alone would transform AI.[9]

He identified models’ weaker generalization relative to people as the fundamental problem. To illustrate the gap between impressive evaluations and practical reliability, he described a coding assistant that fixes one bug, introduces a second, then restores the first when asked to fix the second. He suggested that narrowly focused reinforcement learning and training environments inspired by evaluations might help explain such behavior; these were proposed explanations for an illustrative example.[9]

Sutskever saw value functions as a way to make reinforcement learning more efficient. Rather than waiting until a long sequence of actions produces a final answer, a value function could recognize that an intermediate direction is unpromising and supply an earlier learning signal. He qualified the distinction: anything achievable with a value function could, in his view, be achieved without one, just more slowly.[9]

Sutskever’s November 2025 account qualified SSI’s plan to go directly toward safe superintelligence. He valued freedom from daily market competition, but said long timelines or the benefits of powerful AI reaching the public could cause the plan to change. He also said he had come to place greater importance on incremental deployment, while leaving open whether that shift would affect SSI’s plans. Even under a direct path to superintelligence, he imagined a gradual release.[9]

His envisioned system would continue learning through deployment. He compared it to an eager, superintelligent student who could learn to become a programmer or doctor, rather than a finished mind that already knew every job. Deployment would therefore involve learning and trial and error—a conceptual account of superintelligence, rather than a description of a released SSI system.[9]

Before SSI, Sutskever worked on sequence learning and held research roles at Google Brain and OpenAI. That earlier work provides a concrete example of the training methods and optimization choices behind his later discussion of research and scaling.[10][3][4][9]

Sutskever’s personal career account traces his early work from studying in Toronto’s machine-learning group with Geoffrey Hinton to a Stanford postdoc with Andrew Ng’s group, cofounding DNNresearch, and three years as a research scientist at Google Brain.[10]

In 2014, Sutskever, Oriol Vinyals, and Quoc V. Le coauthored Sequence to Sequence Learning with Neural Networks. Their end-to-end method used a multilayer Long Short-Term Memory network (LSTM) to encode an input sequence into a fixed-dimensional vector, then another deep LSTM to decode the target sequence. The architecture allowed input and output sequences to be learned together.[3]

On WMT’14 English-to-French translation, the authors reported 34.8 BLEU for their LSTM translations, including a penalty for out-of-vocabulary words, compared with 33.3 for a phrase-based statistical machine-translation system. Using the LSTM to rerank that system’s 1,000 hypotheses produced 36.5 BLEU. A consequential optimization choice was to reverse the words in source sentences while preserving target order: the authors found that this introduced shorter dependencies between source and target and made training easier.[3]

OpenAI’s December 11, 2015 founding announcement identified Sutskever as research director. In July 2023, he and Jan Leike jointly authored Introducing Superalignment, framing the question: “How do we ensure AI systems much smarter than humans follow human intent?”[4][5]

Asked in November 2025 about his own research taste, Sutskever emphasized “beauty, simplicity, elegance, correct inspiration from the brain.” He described top-down beliefs about how a system should work as a reason to keep debugging when disappointing experiments might reflect an implementation bug rather than a mistaken idea.[9]