Jakub Pachocki
Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,600 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,600 words
Add missing citations, update stale details, or suggest a clearer explanation.
Jakub Pachocki (born 1991) is a Polish computer scientist and former competitive programmer who has served as the Chief Scientist of OpenAI since May 2024.[^1][^2] He succeeded Ilya Sutskever in the role and is widely regarded as one of the principal technical figures behind several of OpenAI's most consequential systems, including the GPT-4 large language model, the OpenAI Five Dota-playing system, and the reasoning model series beginning with o1 and o3.[^1][^3][^4]
Before joining OpenAI in 2017, Pachocki was an accomplished competitive programmer who represented the University of Warsaw at the ACM International Collegiate Programming Contest (ICPC) World Finals, winning a gold medal at the 2012 World Finals, and won the global championship of Google Code Jam the same year.[^5] He completed his Ph.D. in theoretical computer science at Carnegie Mellon University in 2016 with a dissertation on faster algorithms for high-dimensional convex optimization, before transitioning to applied machine-learning research.[^6]
Pachocki is known for keeping a deliberately low public profile relative to his seniority. Sam Altman, who announced his promotion in May 2024, has publicly described him as "easily one of the greatest minds of our generation,"[^1][^7] and credits Pachocki with co-leading the reinforcement-learning scaling work, GPT-4 pretraining, and the reasoning paradigm that produced o1 and its successors.[^8] In 2025 he was named to the Time 100 Most Influential People in AI list.[^9]
| Born | 1991, Gdańsk, Poland[^10] |
| Nationality | Polish[^10] |
| Education | B.Sc., University of Warsaw; Ph.D., Carnegie Mellon University (2016)[^6][^10] |
| Doctoral advisor | Gary L. Miller[^6] |
| Doctoral thesis | "Graphs and Beyond: Faster Algorithms for High Dimensional Convex Optimization" (2016)[^6] |
| Current role | Chief Scientist, OpenAI (since May 2024)[^1] |
| Previous OpenAI roles | Research Lead, Dota project; Director of Research[^1][^11] |
| Joined OpenAI | 2017[^1] |
| Notable handle | merettm (X / TopCoder)[^5] |
Pachocki was born in 1991 in Gdańsk, Poland.[^10] As a teenager he was a six-time finalist in the Polish Olympiad in Informatics and represented Poland at international competitions in computer science.[^10] In 2009 he won a silver medal at the International Olympiad in Informatics (IOI), finishing 49th in the individual standings.[^5][^10]
Pachocki then enrolled at the University of Warsaw, where he completed an undergraduate degree in computer science.[^1][^12] The Warsaw program, long known for its strength in algorithms and combinatorics, was at the time one of the most successful in the world in competitive programming, and Pachocki quickly became one of its leading students.[^12]
In 2013, after completing his undergraduate studies, Pachocki moved to Pittsburgh to begin a Ph.D. in computer science at Carnegie Mellon University.[^6][^10] He worked under the supervision of theoretical computer scientist Gary L. Miller, with a thesis committee that also included Anupam Gupta, Daniel Sleator, and Shang-Hua Teng of the University of Southern California.[^6] His dissertation, "Graphs and Beyond: Faster Algorithms for High Dimensional Convex Optimization," was conferred in May 2016 and ran to roughly 200 pages.[^6] The work developed nearly-linear-time algorithms for several core problems in convex optimization, including a fast algorithm for the Fermat–Weber problem in robust estimation, what was at the time the fastest algorithm for solving Laplacian linear systems on undirected graphs, and new clustering and oblivious-routing algorithms for directed graphs.[^6] The keywords listed with the thesis were convex optimization, linear-system solvers, and spectral graph theory.[^6]
Before and during graduate school Pachocki accumulated industry experience as a software engineering intern. He interned at Facebook in 2011–2012.[^10] After completing his Ph.D. at CMU, he held postdoctoral appointments at Harvard University and at the Simons Institute for the Theory of Computing at the University of California, Berkeley.[^10][^13] During this period he continued to publish on graph algorithms and spectral methods, including work on resparsification and on optimal lower bounds for sketching and sampling problems.[^14][^15]
Pachocki was one of the most decorated competitive programmers of his generation. His record across major international contests includes:[^5][^10]
His competitive-programming handle is "meret," still used as his X/Twitter handle (@merettm).[^5][^16] By the time he moved into research, he was widely considered one of the strongest contest programmers in the world, a background that several profiles have argued shaped his later instincts for both algorithm design and large-scale empirical AI research.[^4][^8]
Pachocki joined OpenAI in 2017, leaving an academic trajectory in theoretical computer science.[^1][^4] At OpenAI he started as a research lead on the Dota 2 team, the group that would build OpenAI Five, a reinforcement-learning system trained to play the multiplayer online battle arena game Dota 2.[^17][^18] OpenAI Five became, in April 2019, the first AI system to defeat a reigning world champion at an esports game, beating Team OG in a televised match.[^18][^19]
Pachocki is listed among the principal authors of the project's 2019 technical paper, Dota 2 with Large Scale Deep Reinforcement Learning, alongside Greg Brockman, Brooke Chan, and dozens of others.[^19] His specific contributions, as described in the paper and subsequent reporting, centered on the design and scaling of the distributed reinforcement-learning training system, known internally as "Rapid," and on the engineering practices that allowed the team to continually retrain a single agent over roughly ten months while the game and the agent's policy changed.[^4][^19] He worked closely on this with fellow Polish researcher and longtime collaborator Szymon Sidor.[^7]
After OpenAI Five, Pachocki shifted into language-model research as OpenAI's focus moved toward scaling transformer-based systems. He subsequently led the Reasoning Team and the Science of Deep Learning team within OpenAI, and in 2021 was elevated to Director of Research.[^1][^20] In that role he was, per OpenAI's own announcement, the overall lead and the optimization lead for GPT-4, the large multimodal model released in March 2023.[^21] Sam Altman would later credit Pachocki and Szymon Sidor as the people who "led GPT-4 pretraining."[^8]
Public reporting in 2023 indicated that by the middle of that year Pachocki was already helping reorient the company around what would become its reasoning-model program, eventually code-named "Strawberry" inside the company.[^22] He continued in the Director of Research role through the November 2023 turmoil at OpenAI surrounding the brief removal of Sam Altman as CEO.
On May 14, 2024, OpenAI announced that Ilya Sutskever, the company's co-founder and Chief Scientist, would be leaving, and that Pachocki would succeed him as Chief Scientist.[^1][^7] Sam Altman, announcing the change, wrote that Pachocki was "easily one of the greatest minds of our generation," and described him as having previously served as Director of Research, "spearheading the development of GPT‑4 and OpenAI Five" and fundamental research in large-scale reinforcement learning and deep-learning optimization.[^1] Pachocki, in a brief public statement at the time, said that Sutskever "introduced me to the world of deep learning research, and has been a mentor to me, and a great collaborator for many years."[^1]
Shortly after his appointment, Pachocki was named to OpenAI's newly formed Safety and Security Committee, the internal body charged with making safety and security recommendations on critical model releases.[^23]
In September 2024, OpenAI released the o1 series of reasoning models—models trained with large-scale reinforcement learning to perform extended chain-of-thought computation at inference time.[^24][^25] Pachocki has been described in profile pieces in the New York Times and MIT Technology Review as one of the architects of this approach.[^4][^22] Speaking to the New York Times on the launch, he said: "With previous models like ChatGPT, you ask them a question and they immediately start responding. This model can take its time. It can think through the problem—in English—and try to break it down and look for angles in an effort to provide the best answer."[^24]
In December 2024, OpenAI previewed o3, the next reasoning model in the series, which on internal benchmarks outperformed o1 and, as the company noted, exceeded the score Pachocki himself had achieved on a competitive-programming evaluation.[^26]
In 2025 OpenAI's reasoning models continued to set records on programming and mathematics competitions: in July 2025 OpenAI announced that its experimental reasoning system had achieved a gold-medal-level performance on the International Mathematical Olympiad,[^4][^27] and in August 2025 the same family of systems achieved a gold-medal-level score at the IOI 2025, ranking 6th out of 330 human contestants with 533.29 points.[^28] Pachocki has since publicly emphasized chain-of-thought monitorability as a distinctive safety affordance of reasoning systems and co-signed an industry-wide statement on its preservation.[^29]
In an MIT Technology Review profile published July 31, 2025, Pachocki described his role at OpenAI as "responsible for setting the research roadmap and establishing our long-term technical vision," with Chief Research Officer Mark Chen co-leading research on the operational side.[^4] He told the magazine that the field was "probably still…at the very beginning of this reasoning paradigm," that "building smarter-than-human machines is inherently dangerous," and that "alignment problems are now very practically motivated."[^4] Pachocki has publicly asserted that, in his view, deep-learning systems are "less than a decade away from superintelligence," which he defined as systems "smarter than all of us on a large number of critical axes."[^4][^30]
Pachocki's research output spans two clearly distinct phases.
Theoretical computer science (≈2011–2017). During his time at the University of Warsaw, Carnegie Mellon, Harvard, and Simons, Pachocki published a series of papers on graph algorithms, spectral graph theory, and high-dimensional convex optimization, often with collaborators including Gary Miller, Richard Peng, and Jelani Nelson.[^14][^15][^31] His dissertation work delivered new bounds for Laplacian linear-system solvers, geometric median (Fermat–Weber) problems, and clustering, much of it organized around the idea of obtaining algorithms with running time nearly linear in input sparsity.[^6]
Applied AI research at OpenAI (2017–present). Pachocki is named as a contributor on the 2019 Dota 2 with Large Scale Deep Reinforcement Learning paper.[^19] On the GPT-4 technical report and contributions page he is identified as the overall lead and optimization lead for the project.[^21] He is one of the architects of the o1 and o3 reasoning models and an advocate of the test-time-compute scaling paradigm—the idea that increasing the amount of computation a model performs at inference time, particularly through long chains of thought trained with reinforcement learning, yields qualitatively new capabilities in mathematics, science, and code.[^4][^24] More recent OpenAI work he has been associated with, through co-authored papers or technical reports, includes Competitive Programming with Large Reasoning Models and research on chain-of-thought monitorability and the risks of obfuscation.[^32][^33]
Pachocki is unusually low-profile for a research leader of his rank. He maintains an X account (@merettm) on which he posts only sporadically,[^16][^29] and gave essentially no extended public interviews in the years he was leading the Dota and GPT-4 efforts. After his appointment as Chief Scientist, his media presence increased modestly: he gave on-the-record comments to the New York Times on the launch of o1,[^24] sat for the long MIT Technology Review profile alongside Mark Chen in mid-2025,[^4] and appeared on several podcasts, including the OpenAI Podcast,[^34] the "Before AGI" podcast in conversation with Szymon Sidor,[^35] and the a16z Podcast episode "From Vibe Coding to Vibe Researching" in September 2025.[^36]
He was included in the 2025 Time 100 Most Influential People in AI, where he was profiled as one of the principal architects of OpenAI's reasoning paradigm.[^9] Sam Altman has separately written a blog post titled "Jakub and Szymon" celebrating Pachocki and Szymon Sidor as people who had "not been given as much credit" externally as they were due, crediting them with pioneering work on reinforcement-learning scaling, leading GPT-4 pretraining, and contributing to the reasoning breakthroughs.[^8] Altman wrote that Pachocki described Sidor as "indefatigable," and that "OpenAI has not yet thrown a problem at them they have not been able to solve."[^8]
Pachocki is sometimes profiled in Polish-language press as part of a broader story about the success of Polish competitive programmers and computer-science graduates in the global AI industry.[^12][^37]