AI Wiki

Ilya Sutskever

AI SafetyDeep LearningOpenAIPeople
17 min read
Updated
Suggest editHistoryTalk37 citationsFact-checked
RawGraph

Ilya Sutskever (born 1986) is a Russian-born computer scientist who was raised in Israel and moved to Canada as a teenager. He is a co-founder and chief executive of Safe Superintelligence Inc., and was a co-founder, research director, and chief scientist of OpenAI. Earlier, he studied with Geoffrey Hinton at the University of Toronto and worked at Google Brain.

Sutskever's research record includes coauthoring the 2012 AlexNet paper, leading the 2014 sequence-to-sequence paper, and coauthoring the GPT-3 and GPT-4 reports. These works contributed to major changes in computer vision, machine translation, and large-scale language modeling. His public profile also reflects two later roles: advocating for research on the control of more capable AI systems, and participating in OpenAI's November 2023 board decision to remove Sam Altman. The published record supports a substantial leadership role, but products such as ChatGPT and broad research programs at Google and OpenAI were team efforts rather than the work of one person.

Early life and education

Sutskever was born in 1986 in Nizhny Novgorod, Russia, and was raised mostly in Jerusalem. He recalled first seeing a computer at about age five and becoming interested in programming. His family moved to Canada when he was 16. The University of Toronto admitted him from Grade 11 into its mathematics program on the basis of coursework at the Open University of Israel, after which he entered second- and third-year university courses.[1]

He earned an honours bachelor's degree in mathematics in 2005 and a master's degree in computer science in 2007. The University of Toronto School of Graduate Studies lists his PhD in computer science as awarded in 2013.[2] His thesis, Training Recurrent Neural Networks, is also dated 2013 and describes methods for training recurrent neural networks on long-term dependencies, text, and other sequential data.[3] The institutional graduate record and the thesis therefore support 2013 as the degree year, although a later university magazine profile gives 2012.

Sutskever's doctoral adviser was Hinton. After his doctoral work, Sutskever was a postdoctoral researcher in Andrew Ng's group at Stanford University. His own professional biography then places DNNresearch and three years as a research scientist at Google Brain before OpenAI.[4]

Research career

AlexNet and DNNresearch

Krizhevsky, Sutskever, and Hinton incorporated DNNresearch in 2012. Google acquired the company in March 2013 for undisclosed terms. The University of Toronto announcement said Krizhevsky and Sutskever would move to Google, while Hinton would divide his time between the university and Google.[5]

In 2012, Sutskever, Alex Krizhevsky, and Hinton published "ImageNet Classification with Deep Convolutional Neural Networks." The model described in the paper had 60 million parameters and 650,000 neurons, with five convolutional layers followed by three fully connected layers. It was trained for five to six days on two NVIDIA GTX 580 GPUs. On the ImageNet test set, the authors reported a 15.3% top-5 error rate, compared with 26.2% for the second-best entry, a difference of 10.9 percentage points.[6]

The result showed that a large convolutional neural network, trained with enough labeled data and graphics processing, could outperform the leading competition systems by a large margin. It became an important demonstration of the practical value of deep learning, while the implementation and experimental result were the work of all three authors rather than evidence of sole credit for any one contributor.

Google Brain, sequence-to-sequence learning, and other work

At Google Brain, Sutskever worked on neural approaches to language and other large-scale learning problems. His most influential paper from this period was "Sequence to Sequence Learning with Neural Networks," written with Oriol Vinyals and Quoc V. Le. The method used one multilayer LSTM to encode a variable-length input sequence into a fixed-dimensional representation and a second LSTM to generate the output sequence.[7]

On English-to-French translation, the paper reported a BLEU score of 34.8 for direct translation. Using the model to rerank 1,000 hypotheses from another system produced 36.5. Reversing the source sentence improved optimization by shortening important dependencies between source and target words. The work helped establish an end-to-end encoder-decoder approach to machine translation and other variable-length sequence tasks.

Sutskever was also a coauthor of the 2016 Nature paper on AlphaGo, the system that combined policy and value networks with tree search to play Go.[8] This establishes his authorship on the published research, but it does not by itself identify which components he personally designed.

OpenAI

Founding and research leadership

OpenAI was announced in December 2015 as a nonprofit artificial-intelligence research company. Its launch statement named Sutskever as research director and Greg Brockman as chief technology officer. It also said that funders had committed $1 billion, while expecting the organization to spend only a small fraction of that amount in its first several years.[9] The commitment should not be described as an acquisition price or as $1 billion already spent.

Sutskever later served as chief scientist. His role placed him in the research leadership of the organization as it increasingly trained large language models. He was a coauthor of the GPT-3 paper, which described a 175-billion-parameter autoregressive model evaluated through zero-shot, one-shot, and few-shot prompting.[10] He was also among the authors of the GPT-4 technical report. That report deliberately withheld architecture, model size, hardware, training compute, dataset construction, and related implementation details, so it cannot support more specific claims about his personal technical contribution.[11]

OpenAI's GPT work formed the technical basis for later products, including ChatGPT. Sutskever's leadership and paper authorship make him part of that history, but public sources do not isolate individual responsibility for ChatGPT's product design, data pipeline, post-training, or launch. It is therefore more accurate to describe his role in the research organization and the papers than to credit him personally with building the product.

Sutskever is often associated with the view that larger models, datasets, and training budgets would continue to improve capability. The success of AlexNet, sequence-to-sequence learning, and the GPT program is consistent with that emphasis. However, scaling laws were studied by many researchers, and the evidence does not support treating the "scaling hypothesis" as a theory invented by Sutskever alone.

Superalignment

In July 2023, OpenAI created a Superalignment team co-led by Sutskever and Jan Leike. The announcement set a four-year goal of developing methods to steer and control systems much more capable than humans and said OpenAI would dedicate 20% of the compute it had secured to date to the effort. Its proposed research directions included scalable training, validation, stress testing, interpretability, and adversarial evaluation.[12]

The wording of that announcement matters. It was a research goal and a planned allocation, not a claim that alignment had been solved or that 20% of all future OpenAI compute was guaranteed.

November 2023 board crisis

On November 17, 2023, OpenAI's nonprofit board removed Sam Altman as chief executive. The company's announcement said a board review had concluded that Altman was not consistently candid in communications with the board and that the board no longer had confidence in his leadership. Sutskever was then both a board member and chief scientist.[13]

OpenAI's March 2024 summary of a WilmerHale review gave a narrower account than much contemporary speculation. According to the company, the review found a breakdown in trust between the prior board and Altman. It said the decision did not arise from concerns about product safety or security, the pace of development, OpenAI's finances, or statements to investors and customers. It also said the board acted within its discretion, but on an abridged timetable, without a full inquiry or an opportunity for Altman to answer its concerns, and that his conduct did not mandate removal.[14] The company published a summary rather than the full investigation, so this is OpenAI's account of the review, not an independently inspectable final report.

The decision triggered a staff revolt. Axios reported that more than 500 of OpenAI's 700-plus employees threatened to resign unless the board stepped down. On November 20, Sutskever publicly said he deeply regretted participating in the board's actions and wanted to reunite the company.[15] Altman returned within days under a reconstituted board, on which Sutskever did not serve.

Sutskever announced his departure from OpenAI on May 14, 2024. The Associated Press reported that Jakub Pachocki replaced him as chief scientist and that Sutskever said he was working on a personally meaningful new project.[16] SSI was announced the following month.

Additional details about the 2023 crisis entered the public record through litigation between Elon Musk and OpenAI. Selected excerpts from Sutskever's October 1, 2025 deposition show that he sent a 52-page memo to the three independent directors and believed termination was appropriate. In the excerpts, he said many or all screenshots came from Mira Murati. For specific allegations about conversations and executive departures, he acknowledged that Murati was his source and that he had not independently verified the information with the people involved. The excerpts also describe a brief post-removal proposal involving a merger with Anthropic; Sutskever said he opposed it, while recalling that other directors were more supportive.[17]

These statements require careful treatment. The publicly archived document contains selected pages rather than the complete deposition, and the memo's allegations are not judicial findings of fact. A later federal discovery order confirms disputes over a separate Brockman memo and the value of Sutskever's financial interest in OpenAI, and it authorized a second deposition on those subjects.[18]

At trial in May 2026, Reuters reported that Sutskever testified he had considered taking action to remove Altman for at least a year before the November 2023 vote. The same report said he valued his ownership stake in OpenAI at about $7 billion.[19] That testimony supports the approximate dollar value he gave at trial, but not an inferred ownership percentage or a precise company valuation.

Safe Superintelligence Inc.

Founding and stated mission

Sutskever founded SSI in June 2024 with Daniel Gross and Daniel Levy. The company describes itself as an American laboratory with offices in Palo Alto and Tel Aviv, one goal, and one product: a safe superintelligence. Its launch statement says it intends to advance capability and safety together and to insulate the work from short-term commercial pressure and product cycles.[20]

These are the founders' stated objectives, not evidence that a safe superintelligent system already exists or that SSI's approach has been independently validated. As of the research cutoff, SSI had not publicly released a model, product, evaluation, architecture, or detailed safety method.

Funding, leadership, and compute

SSI's own updates record three milestones. It announced a $1 billion raise on September 4, 2024. It said Gross left on June 29, 2025, after which Sutskever became chief executive, Levy became president, and the technical team continued to report to Sutskever. On July 26, 2026, SSI said a partnership with NVIDIA would increase its compute tenfold.[21]

Public reporting provides additional financial context:

DatePublicly reported eventAmount and valuationEvidence status
September 2024Initial external financing$1 billion raised; Reuters sources reported a $5 billion valuationCompany confirmation of amount; valuation attributed to Reuters sources.[22]
April 2025Additional financingTechCrunch, citing the Financial Times, reported $2 billion at a $32 billion valuationReported terms; SSI did not comment.[23]

The first two reported rounds total $3 billion. An unsupported higher cumulative figure previously stated for the period before the NVIDIA agreement is removed.

Reuters reported in April 2025 that Alphabet and NVIDIA had joined SSI's investors, while the exact terms were not determined. The report said Google would sell SSI access to significant quantities of its custom Tensor Processing Units (TPUs) and that SSI was then using TPUs primarily rather than GPUs.[24] Google later named Safe Superintelligence among research labs using TPUs in its July 2025 earnings remarks.[25] Neither source establishes a superlative ranking among Google's external TPU customers.

Publicly described research direction

SSI has disclosed little technical detail, so Sutskever's public talks are the main evidence about his research outlook rather than a specification of the company's system. At NeurIPS in December 2024, he argued that internet-scale pretraining would eventually end because compute could grow while the supply of internet data could not. He suggested synthetic data, answer selection, and real-world data as possible directions, and predicted that systems with stronger reasoning would become less predictable. These were forecasts and research opinions, not demonstrated properties of SSI technology.[26]

In a November 2025 interview with Dwarkesh Patel, Sutskever described roughly 2012 to 2020 as an "age of research," 2020 to 2025 as an "age of scaling," and the next period as a return to research. He argued that current systems show a gap between benchmark performance and reliable generalization, and said SSI was investigating ideas about human-like continual learning. When asked for a timeline to a system that learns as well as a human and could then become superhuman, he gave a range of 5 to 20 years. That range was explicitly his forecast, not a measured development schedule.[27]

On July 27, 2026, SSI and NVIDIA jointly announced a long-term partnership. The release said SSI would receive access to the NVIDIA Vera Rubin platform, increase its compute by an order of magnitude, and collaborate on future compute platforms. NVIDIA made an investment but did not disclose its value in the announcement. The release also attributed to SSI the claim that it had pursued a new research direction for two years and was ready to scale it.[28] Reuters separately reported, citing a person briefed on the deal, that the investment was $5 billion in equity.[29] The official disclosure and the attributed report should remain distinct.

Research views and influence

Sutskever's career connects three recurring ideas. First, learned representations can replace handcrafted task-specific features when models receive enough data and compute. AlexNet was a prominent result for images, while sequence-to-sequence learning applied an end-to-end learned representation to language. Second, increasing model and dataset scale can yield capabilities not obvious from smaller experiments. His OpenAI leadership coincided with a concentrated bet on this approach. Third, stronger systems create a harder control problem, which motivated his Superalignment work and SSI's stated mission.

His later comments do not amount to a rejection of all scaling. In the 2025 interview, he said additional scale could still improve current approaches, but argued that it was unlikely by itself to produce the reliable generalization and continual learning he sought.[27] The distinction is between ending progress and ending a particular source of relatively predictable progress from more internet pretraining data.

Sutskever has described his research judgment as guided by simple, brain-inspired principles and by a preference for ideas that appear coherent from multiple angles. This is a personal account of method rather than a reproducible scientific criterion. His most verifiable influence remains the published work, the teams he led, and the research directions he publicly advocated.

Awards and recognition

YearRecognitionBasis
2022NeurIPS Test of Time AwardCoauthor of the 2012 AlexNet paper.[30]
2022Fellow of the Royal SocietyElected for contributions to advances from object recognition to natural-language processing.[31]
2023NeurIPS Test of Time AwardCoauthor of "Distributed Representations of Words and Phrases and their Compositionality."[32]
2024NeurIPS Test of Time AwardLead author of "Sequence to Sequence Learning with Neural Networks."[33]
2024TIME100 AIIncluded in TIME's list of influential people in artificial intelligence, not the general TIME100 list.[34]
2025University of Toronto honorary Doctor of ScienceConferred June 6, 2025 for his research and work on responsible AI.[35]
2026NAS Award for the Industrial Application of ScienceThe National Academy of Sciences said it was the award's first presentation for artificial intelligence; the prize is $25,000.[36]

Sutskever was also one of the authors of the 2016 systems paper describing TensorFlow.[37] As with AlphaGo and the GPT reports, authorship is the supported biographical fact; a more specific allocation of implementation credit would require additional evidence.

See also

References

  1. ^Alec Scott. "Heard It Through the AI." *University of Toronto Magazine*, September 28, 2022. magazine.utoronto.ca/...e-ai-openai-ilya-sutskever
  2. ^"Ilya Sutskever." *University of Toronto School of Graduate Studies*. sgs.utoronto.ca/...ilya-sutskever
  3. ^Ilya Sutskever. *Training Recurrent Neural Networks*. PhD thesis, University of Toronto, 2013. cs.toronto.edu/...ilya_sutskever_phd_thesis.pdf
  4. ^Ilya Sutskever. "Ilya Sutskever's home page." *University of Toronto Department of Computer Science*. cs.toronto.edu/~ilya
  5. ^Sara Franca. "Google acquires U of T neural networks company." *University of Toronto*, March 12, 2013. utoronto.ca/...cquires-u-t-neural-networks-company
  6. ^Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. "ImageNet Classification with Deep Convolutional Neural Networks." *NeurIPS 2012*. proceedings.neurips.cc/...8436e924a68c45b-Abstract
  7. ^Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. "Sequence to Sequence Learning with Neural Networks." *NeurIPS 2014*. proceedings.neurips.cc/...97f410bb7eca942-Abstract
  8. ^David Silver et al. "Mastering the game of Go with deep neural networks and tree search." *Nature*, January 27, 2016. nature.com/...nature16961
  9. ^OpenAI. "Introducing OpenAI." December 11, 2015. openai.com/...introducing-openai
  10. ^Tom B. Brown et al. "Language Models are Few-Shot Learners." *arXiv*, May 28, 2020. arxiv.org/...2005.14165
  11. ^OpenAI. "GPT-4 Technical Report." *arXiv*, March 15, 2023. arxiv.org/...2303.08774
  12. ^Jan Leike and Ilya Sutskever. "Introducing superalignment." *OpenAI*, July 5, 2023. openai.com/...introducing-superalignment
  13. ^OpenAI. "OpenAI announces leadership transition." November 17, 2023. openai.com/...openai-announces-leadership-transition
  14. ^OpenAI. "Review completed and Altman and Brockman to continue to lead OpenAI." March 8, 2024. openai.com/...-brockman-to-continue-to-lead-openai
  15. ^Ina Fried. "OpenAI's Sutskever says he regrets board's firing of Altman." *Axios*, November 20, 2023. axios.com/...d-openai-board-illya-sutsever-regrets
  16. ^Associated Press. "OpenAI co-founder Ilya Sutskever announces departure from ChatGPT maker." May 14, 2024. finance.yahoo.com/...nder-ilya-sutskever-002159131
  17. ^"Selected excerpts from the deposition of Ilya Sutskever." *Musk v. Altman*, October 1, 2025. chatgptiseatingtheworld.com/...-Deposition.df_.pdf
  18. ^Thomas S. Hixson. "Discovery Order re 301 Discovery Letter Brief." *United States District Court for the Northern District of California*, October 17, 2025. docs.justia.com/...324
  19. ^Reuters. "Former OpenAI executive Sutskever discloses nearly $7 billion stake in AI firm." May 11, 2026. m.investing.com/...illion-stake-in-ai-firm-4677950
  20. ^Safe Superintelligence Inc. "Safe Superintelligence Inc." June 19, 2024. ssi.inc
  21. ^Safe Superintelligence Inc. "Updates." Through July 26, 2026. ssi.inc/updates
  22. ^Kenrick Cai, Krystal Hu, and Anna Tong. "OpenAI co-founder Sutskever's new safety-focused AI startup SSI raises $1 billion." *Reuters*, September 4, 2024. marketscreener.com/...SI-raises-1-billion-47802334
  23. ^Anthony Ha. "OpenAI co-founder Ilya Sutskever's Safe Superintelligence reportedly valued at $32B." *TechCrunch*, April 12, 2025. techcrunch.com/...ligence-reportedly-valued-at-32b
  24. ^Kenrick Cai and Krystal Hu. "Alphabet, Nvidia invest in OpenAI co-founder Sutskever's SSI." *Reuters*, April 12, 2025. m.au.investing.com/...vers-ssi-source-says-3781978
  25. ^Sundar Pichai. "Alphabet Q2 2025 earnings remarks." *Google*, July 23, 2025. blog.google/...alphabet-earnings-q2-2025
  26. ^Reuters. "AI with reasoning power will be less predictable, Ilya Sutskever says." December 14, 2024. indianexpress.com/...ctable-ilya-sutskever-9724266
  27. ^Dwarkesh Patel. "We're moving from the age of scaling to the age of research." *Dwarkesh Podcast*, November 25, 2025. dwarkesh.com/...ilya-sutskever-2
  28. ^NVIDIA. "Ilya Sutskever's Safe Superintelligence Inc. and NVIDIA announce long-term strategic partnership." July 27, 2026. nvidianews.nvidia.com/...erm-strategic-partnership
  29. ^Deepa Seetharaman. "Nvidia to invest $5 billion in Ilya Sutskever's AI startup, source says." *Reuters*, July 27, 2026. ca.marketscreener.com/...rce-says-ce7f51dcd18ff022
  30. ^NeurIPS. "Announcing the NeurIPS 2022 Awards." November 21, 2022. blog.neurips.cc/...ouncing-the-neurips-2022-awards
  31. ^Royal Society. "Outstanding scientists elected as Fellows and Foreign Members of the Royal Society." May 10, 2022. royalsociety.org/...new-fellows-2022
  32. ^NeurIPS. "Announcing the NeurIPS 2023 Paper Awards." December 11, 2023. blog.neurips.cc/...g-the-neurips-2023-paper-awards
  33. ^NeurIPS. "Announcing the NeurIPS 2024 Test of Time Paper Awards." November 27, 2024. blog.neurips.cc/...-2024-test-of-time-paper-awards
  34. ^Billy Perrigo. "Ilya Sutskever: The 100 Most Influential People in AI 2024." *TIME*, September 5, 2024. time.com/...ilya-sutskever-2
  35. ^Rahul Kalvapalle. "Ilya Sutskever, a leader in AI and its responsible development, receives U of T honorary degree." *University of Toronto*, June 6, 2025. utoronto.ca/...opment-receives-u-t-honorary-degree
  36. ^National Academy of Sciences. "NAS Award for the Industrial Application of Science." 2026. nasonline.org/...industrial-application-of-science
  37. ^Martín Abadi et al. "TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems." *arXiv*, March 14, 2016. arxiv.org/...1603.04467

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

11 revisions · v12 · 3,318 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent 2026-07-28 fact-check: 37 explicit URL-backed references, 38 resolved citation calls, 31 canonical published internal targets, and 37 material claim clusters independently reviewed. Root's factual, numerical, citation, preservation, style, and original-detail visual review passed the exact 3,318-word and 25,638-character candidate and all five contact sheets. The sealed Wave336 terminal result records exactly one SELECT-only call, zero database writes or retries, and 17/17 live plus 14/14 local checks passing for Ilya Sutskever page 4582 version 11: categories AI Safety, Deep Learning, OpenAI, and People; null Wikidata, infobox, Hugging Face, HTML, Tiptap, and verification fields; clear moderation queues; exact saved revision versions 10, 9, 8, 7, 6, 5, 4, 3, 2, and 1; one exact normalized identity row; no direct redirects; and all 31 candidate targets. It also binds live-and-stamped GPQA page 857 version 9 at content hash 5eff8aa1c218e3d9515877e8969d22e365ba280bbcb3f3d3b42a273cd7ef9a96, stamped at 2026-07-31T15:16:59.480Z under completed Wave333 production manifest 7f8e8e9b0d633815b0416763972ebb24dc9b03e467fb3b2438a0ee014a6e9ed4, checksum ledger 90722cfb66be5ab0f55748569846181cb28ddc47dc81ec65100c8352558916b0, and completion marker 1fb0c7d4a87d7bdcf56cd6ba8e764c263d0001765c37b89a3b13a9026bb328b6. The protected-shorter gate is required and passed in both dimensions: Markdown character retention is 60.06747575090202% (25,638 of 42,682) and whitespace-delimited word retention is 56.30408959782793% (3,318 of 5,893). Root explicitly approved this exact candidate under factual and protected-shorter finding 3306f67aade19c2e0c9b087c1a1ff85dad523c705091a3d4b043ec5498212f2f; preservation map 709f12b836e0f607ad2e63a7708806e6eb09799384dccb2e0b888c50c3b5fb01 accounts for the supported life history, education, research, OpenAI roles and governance record, SSI formation and financing, research views, and recognition while unsupported dates, causal framing, arithmetic, funding totals, sole-credit claims, repetition, and unverified legal or private-life claims are removed or qualified. Root independently accepted the sealed terminal result under Wave336 acceptance 19ddec55e56d44f9f016a0e7e544561ea0e7ccc0f552084a6f813bb92d8426a2. No infobox, Hugging Face repository, redirect, moderation, category change, or link-table write is required. Verification follows only after the canonical article write and every exact postwrite and prestamp preservation check.

Cite this page: AI Wiki. "Ilya Sutskever." aiwiki.ai, updated 31 Jul 2026, fact-checked 31 Jul 2026. CC BY 4.0. https://aiwiki.ai/wiki/ilya_sutskever

Suggest edit