Jeff Dean
Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 5,653 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
v2 · 5,653 words
Add missing citations, update stale details, or suggest a clearer explanation.
Jeffrey Adgate "Jeff" Dean (born 23 July 1968) is an American computer scientist who serves as Google Senior Fellow and Chief Scientist of Google DeepMind and Google Research. He joined Google in August 1999 as one of its earliest engineers and went on to architect or co-architect a long sequence of systems that shaped the modern internet, including the indexing pipeline behind Google Search, the MapReduce programming model, the Bigtable distributed storage system, the Spanner globally distributed database, Protocol Buffers, the TensorFlow machine learning framework, the Tensor Processing Unit hardware program, the Pathways ML system, the PaLM language model, and the Gemini family of multimodal models. Dean was a co-founder of Google Brain in 2011 with Andrew Ng and Greg Corrado, led Google AI from April 2018, and from April 2023 has served as Chief Scientist of Google DeepMind and Google Research, co-leading the Gemini program with Oriol Vinyals and, from August 2024, Noam Shazeer [^1][^2][^3]. He was elected to the U.S. National Academy of Engineering in 2009, made an ACM Fellow the same year, awarded the IEEE John von Neumann Medal in 2021, and elected to the American Academy of Arts and Sciences in 2016 [^1][^4][^5].
| Field | Detail |
|---|---|
| Born | 23 July 1968 |
| Nationality | American |
| Alma mater | University of Minnesota (B.S., 1990); University of Washington (Ph.D., 1996) |
| Doctoral advisor | Craig Chambers |
| Employer | Google (since August 1999) |
| Current title | Google Senior Fellow; Chief Scientist, Google DeepMind and Google Research |
| Notable systems | MapReduce, Bigtable, Spanner, Protocol Buffers, TensorFlow, TPU, Pathways, PaLM, Gemini |
| Notable awards | ACM Fellow (2009), NAE (2009), ACM-Infosys Foundation Award (2012), Mark Weiser Award (2012), American Academy of Arts and Sciences (2016), IEEE John von Neumann Medal (2021), ACM SIGMOD Systems Award for Spanner (2025) |
Dean was born on 23 July 1968. He completed a Bachelor of Science in computer science and economics, summa cum laude, at the University of Minnesota in 1990. His undergraduate honors thesis described parallel implementations of neural network training, work that he would later say presaged his interest in scaling machine learning [^1][^4].
He earned his Ph.D. in computer science from the University of Washington in 1996 under the supervision of Craig Chambers. His dissertation, "Whole-Program Optimization of Object-Oriented Languages," introduced techniques for static analysis and optimization of programs written in object-oriented languages, including dynamic dispatch elimination, class hierarchy analysis, and inlining transformations. The work was carried out in the context of the Cecil compiler and influenced later compiler research on language implementations such as Java and Self [^5][^6]. While at Washington he also collaborated with David Grove and others on papers about call graph construction and selective specialization that have continued to be cited in compiler textbooks.
Before and during graduate school Dean held research positions that shaped his later interests. As an undergraduate he wrote statistical software for the World Health Organization's Global Programme on AIDS at the U.S. Centers for Disease Control during 1988 and 1989, producing tools used in epidemiological projections of the HIV/AIDS pandemic [^4][^7]. He once described those summers as his first experience writing code that other people depended on, and credited the work for orienting him toward systems whose performance directly mattered to users.
From 1996 to 1999, after completing his doctorate, Dean worked at the DEC Western Research Lab in Palo Alto, California. The lab, then a leading industrial research outpost, had assembled engineers including Sanjay Ghemawat, Krste Asanovic, Mike Burrows, and Luiz Andre Barroso, several of whom would later follow Dean to Google. At WRL, Dean worked on profiling tools, hardware performance analysis, and a Java implementation. He co-authored the influential 1997 paper "ProfileMe: Hardware Support for Instruction-Level Profiling on Out-of-Order Processors," which described mechanisms for sampling individual instruction traces in modern superscalar CPUs, and the 1997 "Continuous Profiling: Where Have All the Cycles Gone?" paper on the DCPI system [^8][^9]. Both became reference points in the systems performance literature.
Dean and Ghemawat began their collaboration at WRL, and their habit of pair-programming on infrastructure problems would carry over to Google. The closure of Compaq's research labs in 2001 led several DEC alumni to join Google in the early 2000s, but Dean and Ghemawat had already moved over by 1999.
Dean joined Google in August 1999, when the company occupied a building above a bicycle shop in Palo Alto and had roughly twenty employees. His first projects involved Google's advertising system and the search indexing pipeline. He rewrote the early ads serving infrastructure to handle higher query rates and built parts of the crawling and indexing system that allowed Google's index to grow from millions to billions of documents [^10][^11].
In the early 2000s Dean and Ghemawat worked on the systems that became Google's signature: the Google File System (GFS) provided distributed storage; MapReduce supplied a programming model for batch computation; and Bigtable offered a structured storage abstraction over GFS. These three systems formed the substrate on which much of Google's product line ran during the decade. Dean later said that one of the recurring patterns of his early career was discovering that whatever he and his colleagues had built for one product turned out to be useful for many others, which led to the consistent push to build infrastructure as reusable components.
The paper "MapReduce: Simplified Data Processing on Large Clusters," by Jeffrey Dean and Sanjay Ghemawat, was presented at the Sixth Symposium on Operating System Design and Implementation (OSDI '04) in San Francisco in December 2004 [^12]. MapReduce expressed parallel computations as a sequence of map operations followed by reduce operations, hiding the details of distribution, fault tolerance, and load balancing inside a runtime library. The paper showed that the abstraction was sufficient for indexing, log analysis, machine translation training data preparation, and many other Google workloads. Within a few years it had inspired the open source Apache Hadoop project and a generation of big data tools.
In 2006 Dean co-authored "Bigtable: A Distributed Storage System for Structured Data" with Fay Chang, Sanjay Ghemawat, Wilson Hsieh, Deborah Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert Gruber, presented at OSDI '06 [^13]. Bigtable provided a sparse, distributed, multi-dimensional sorted map running over GFS, with applications across Google including web indexing, Google Earth, Google Analytics, and personalized search. Bigtable's data model influenced Apache HBase, Apache Cassandra, and many other wide-column stores.
Protocol Buffers, a language-neutral mechanism for serializing structured data, was developed inside Google starting in 2001 and released as open source in 2008. Dean and Ghemawat were among the original designers, and Protocol Buffers became Google's default format for inter-process communication, configuration data, and storage [^14]. The system later inspired Apache Thrift and gRPC and is one of the most widely deployed serialization formats in industry.
Spanner, presented at OSDI '12 in a paper authored by James C. Corbett, Dean, and a long list of co-authors, extended Bigtable into a globally distributed, externally consistent transactional database [^15]. Its TrueTime API used GPS receivers and atomic clocks at every datacenter to bound clock uncertainty, enabling externally consistent transactions across continents. Spanner became the storage backbone for Google's advertising platform, Gmail, Google Photos, and many other services, and was later commercialized as Cloud Spanner. In June 2025, Spanner was awarded the ACM SIGMOD Systems Award, with Dean listed among more than thirty named contributors, in recognition of the system's lasting impact on relational data management at global scale [^16][^17].
Through the 2000s Dean and his collaborators contributed to many other systems: AdSense's contextual targeting infrastructure, the AdWords serving system, the news indexing pipeline, query freshness improvements, multi-datacenter sharding for search, the LevelDB key-value store (open sourced in 2011, with Sanjay Ghemawat as primary author and Dean as co-author), and a suite of tools to make distributed debugging tractable. Dean's 2009 talk "Numbers Everyone Should Know," which gave order-of-magnitude latencies for memory access, disk seeks, network round trips, and other operations, became a widely cited reference for systems engineers reasoning about performance budgets [^18]. He also co-authored with Luiz Andre Barroso the 2013 Communications of the ACM article "The Tail at Scale," which framed how latency outliers dominate aggregate response times in large fanout systems and became a standard reference in the field [^19].
In 2011 Dean became one of the three co-founders of Google Brain, alongside Andrew Ng and Greg Corrado. The project began as a part-time research collaboration between Dean, Corrado, and Ng inside Google X. The team's premise was that Google's distributed systems expertise and access to data made it well placed to scale up neural networks at the time when deep learning was beginning to deliver state-of-the-art results in speech recognition and computer vision [^20][^21].
Dean and his collaborators built DistBelief, Google's first large-scale distributed deep learning system, and used it to train neural networks with up to a billion parameters across thousands of machines. The 2012 paper "Building High-level Features Using Large Scale Unsupervised Learning," by Quoc Le, Marc'Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg Corrado, Jeff Dean, and Andrew Ng, became known as the "cat neuron" paper because the network learned, without labels, a feature that selectively responded to images of cats [^22]. The result drew wide press attention and contributed to a surge of investment in deep learning at Google and elsewhere.
DistBelief was used to train models that improved Google's voice recognition and image search, and over the next two years Brain expanded into language modeling, sequence-to-sequence learning, and reinforcement learning. The team's growth eventually led to the design of a successor framework that would become TensorFlow. Brain was the home of much of the early work on the Transformer architecture, word2vec, and large-scale neural machine translation deployed in Google products.
In 2015 Dean co-authored, with Geoffrey Hinton and Oriol Vinyals, the paper "Distilling the Knowledge in a Neural Network," which introduced what is now called knowledge distillation, a technique for transferring the behavior of a large "teacher" model into a smaller "student" model by training on the teacher's softened output distribution [^23]. The paper has become a foundational reference for model compression and is cited in much of the subsequent literature on efficient deployment of large models.
In 2017 Dean was a co-author on "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer," with Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, and Geoffrey Hinton [^24]. The paper described a layer that routed each input to a small number of expert subnetworks selected from thousands, allowing models with hundreds of billions of parameters to be trained at the cost of evaluating only a small fraction per example. Mixture of experts routing has since become a standard ingredient in many of the largest language models, including Google's own Gemini and several open-weight models.
TensorFlow, the successor to DistBelief, was released as open source in November 2015 under an Apache 2.0 license. Dean was one of the co-leads of its development and co-authored the 2016 paper "TensorFlow: A System for Large-Scale Machine Learning" with a long list of Brain colleagues [^25]. TensorFlow became one of the most widely used machine learning frameworks of the second half of the 2010s, with bindings in many languages and ports to mobile and embedded targets. Although PyTorch gained ground in research from 2018 onward, TensorFlow remained dominant in industrial deployments for years and influenced the design of nearly every subsequent ML framework.
In parallel, Dean worked closely with Norm Jouppi and Google's hardware team on the Tensor Processing Unit, Google's family of custom AI accelerators. The first TPU was deployed inside Google starting in 2015 and described publicly in the 2017 paper "In-Datacenter Performance Analysis of a Tensor Processing Unit," which argued that domain-specific hardware could deliver order-of-magnitude improvements in performance per watt for inference workloads [^26]. Subsequent TPU generations expanded into training and into multi-pod systems with thousands of chips connected over high-bandwidth interconnects. The sixth-generation TPU, code-named Trillium (TPU v6e), was announced at Google I/O in May 2024 and reached general availability in late 2024, with peak compute per chip roughly 4.7 times that of TPU v5e [^27]. The seventh generation, Ironwood (TPU v7), was unveiled at Google Cloud Next in April 2025 as Google's first TPU designed explicitly for inference at massive scale, with 192 GB of high-bandwidth memory per chip and 9,216-chip pods delivering 42.5 exaflops of aggregate compute [^28]. Dean has been a public advocate for the co-design of models, software, and silicon, and the TPU program is one of the clearest examples of that approach.
In April 2018 Google reorganized its research and AI efforts, naming Dean head of Google AI, a unit that combined Google Brain, the Google Research engineering teams, and several adjacent groups. He retained his role as a Google Senior Fellow [^29]. From this position he oversaw the Brain team's expansion into multimodal models, robotics, healthcare, and AI for science.
In October 2021 Dean announced Pathways, a new architecture for next-generation AI systems intended to support sparsity, multimodality, and many tasks within a single model. The accompanying technical paper, "Pathways: Asynchronous Distributed Dataflow for ML," was published at MLSys 2022 by Paul Barham, Aakanksha Chowdhery, Dean, Sanjay Ghemawat, Steven Hand, Daniel Hurt, Michael Isard, and others [^30]. Pathways introduced a runtime that scheduled large training jobs across thousands of TPU chips with finer-grained control than earlier frameworks, and it became the substrate on which subsequent Google flagship models were trained.
PaLM, the Pathways Language Model, was unveiled in April 2022 in the paper "PaLM: Scaling Language Modeling with Pathways," by Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, and many others including Dean [^31]. PaLM was a 540-billion parameter dense decoder-only Transformer trained on 6,144 TPU v4 chips, and it set new state-of-the-art scores on a range of language benchmarks, including BIG-bench reasoning tasks. PaLM was followed by PaLM 2 in May 2023 and by the medical PaLM derivatives Med-PaLM and Med-PaLM 2.
In April 2023 Google announced the merger of Google Brain with DeepMind into a single division, Google DeepMind, led by Demis Hassabis [^32]. Dean stepped back from his Google AI head role and was named Chief Scientist of Google DeepMind and Google Research, a position in which he reports directly to CEO Sundar Pichai and oversees AI research across the merged organization. Dean was reported to have been one of the internal advocates for unifying Google's two large AI research groups behind a single roadmap, with Gemini as its centerpiece.
Dean co-led the Gemini effort with Oriol Vinyals. The Gemini 1.0 technical report, released on 6 December 2023, listed Dean among the senior contributors and described a family of natively multimodal models trained jointly on text, images, audio, and video [^33]. Gemini Ultra reported state-of-the-art results on many academic benchmarks, including the MMLU exam-style test where it became the first model to exceed a 90% human-expert-style threshold. Gemini 1.5, released in February 2024, introduced a long-context architecture that handled inputs of up to one million tokens, later extended to two million [^34]. Gemini 2.0, announced in December 2024, focused on agentic capabilities, native tool use, and lower latency. Gemini 2.5 Pro, released in March 2025, introduced built-in extended reasoning ("thinking") and topped multiple coding and reasoning leaderboards on release [^35]. Gemini 3 was announced on 18 November 2025 by Sundar Pichai and Demis Hassabis and rolled out simultaneously across Google Search AI Mode, the Gemini consumer app, and Vertex AI, with reported state-of-the-art scores on multimodal and reasoning benchmarks including 1501 Elo on LMArena and 91.9% on GPQA Diamond [^36][^37].
In August 2024, after Noam Shazeer rejoined Google from Character.AI, Shazeer was named co-technical lead of Gemini alongside Dean and Vinyals, broadening the triumvirate steering subsequent generations of the model [^38]. Dean has separately reported that he proposed the "Gemini" name itself, framing the project as "twins coming together" from the Brain and DeepMind lineages [^39].
Dean has written publicly about the role of AI in scientific discovery, citing collaborations with DeepMind's biology efforts including AlphaFold, AlphaProteo, and AlphaMissense, as well as Google Research projects in weather forecasting (GraphCast, GenCast), chip design (AlphaChip), and mathematical reasoning (AlphaProof, AlphaGeometry). In May 2025, DeepMind announced AlphaEvolve, a Gemini-powered evolutionary coding agent that pairs language models with automated evaluators to discover improved algorithms; Dean publicly highlighted that AlphaEvolve had "proposed a circuit design so counterintuitive yet efficient that it was integrated directly into the silicon of our next-generation TPUs" [^40][^41].
While leading Gemini, Dean has continued to publish on systems and on energy efficiency. His talk "Important Trends in AI: How Did We Get Here, What Can We Do Now and How Can We Shape AI's Future?", a refresh of his "Exciting Trends in Machine Learning" lecture, was delivered at the NeurIPS 2024 ML for Systems workshop in Vancouver, as a Distinguished Lecture at ETH Zürich in April 2025, at the Stanford AI Club in November 2025, and at multiple universities through 2026 [^42][^43][^44]. The talks describe a roughly thousand-fold gain in performance per dollar and per joule for Google's machine learning workloads since 2014, attributed to a combination of TPU generations, sparsity, model distillation, quantization, and software efficiencies, and argue that further multi-order-of-magnitude gains are achievable through hardware-software co-design. In 2025 and 2026 he was also a co-author on "Decoupled DiLoCo," a Google DeepMind research paper on resilient, asynchronous distributed pretraining that allows training across heterogeneous hardware with substantially reduced bandwidth requirements [^45][^46].
| Year | Role |
|---|---|
| 1988 to 1989 | Programmer, World Health Organization Global Programme on AIDS, U.S. CDC |
| 1990 | B.S., University of Minnesota |
| 1996 | Ph.D., University of Washington |
| 1996 to 1999 | Research engineer, DEC Western Research Laboratory |
| 1999 | Joined Google |
| 2009 | Named Google Senior Fellow |
| 2011 | Co-founded Google Brain with Andrew Ng and Greg Corrado |
| April 2018 | Named head of Google AI |
| April 2023 | Named Chief Scientist of Google DeepMind and Google Research; co-lead of Gemini with Oriol Vinyals |
| August 2024 | Noam Shazeer named third Gemini co-technical lead alongside Dean and Vinyals |
| June 2025 | Spanner team, including Dean, awarded ACM SIGMOD Systems Award |
| November 2025 | Gemini 3 launched across Google Search, Gemini app, and Vertex AI |
| Year | Paper | Venue |
|---|---|---|
| 1997 | ProfileMe: Hardware Support for Instruction-Level Profiling | MICRO |
| 1997 | Continuous Profiling: Where Have All the Cycles Gone? | SOSP |
| 2004 | MapReduce: Simplified Data Processing on Large Clusters | OSDI |
| 2006 | Bigtable: A Distributed Storage System for Structured Data | OSDI |
| 2008 | MapReduce: Simplified Data Processing on Large Clusters | CACM (revised) |
| 2012 | Building High-level Features Using Large Scale Unsupervised Learning | ICML |
| 2012 | Spanner: Google's Globally-Distributed Database | OSDI |
| 2012 | Large Scale Distributed Deep Networks (DistBelief) | NeurIPS |
| 2013 | Efficient Estimation of Word Representations in Vector Space (word2vec) | ICLR Workshop |
| 2013 | The Tail at Scale | CACM |
| 2015 | Distilling the Knowledge in a Neural Network | NeurIPS Deep Learning Workshop |
| 2016 | TensorFlow: A System for Large-Scale Machine Learning | OSDI |
| 2017 | Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer | ICLR |
| 2017 | In-Datacenter Performance Analysis of a Tensor Processing Unit | ISCA |
| 2018 | A Guide to Deep Learning in Healthcare | Nature Medicine |
| 2022 | Pathways: Asynchronous Distributed Dataflow for ML | MLSys |
| 2022 | PaLM: Scaling Language Modeling with Pathways | JMLR |
| 2023 | Gemini: A Family of Highly Capable Multimodal Models | Technical report |
| 2024 | Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens of Context | Technical report |
| 2025 | Decoupled DiLoCo for Resilient Distributed Pre-training | Google DeepMind / arXiv |
| Year | Award |
|---|---|
| 2007 | Fellow, ACM SIGOPS Hall of Fame (with co-authors), for various foundational papers (multiple inductions over subsequent years) |
| 2009 | Member, U.S. National Academy of Engineering |
| 2009 | ACM Fellow |
| 2009 | Named Google Senior Fellow |
| 2012 | ACM-Infosys Foundation Award (jointly with Sanjay Ghemawat); later renamed the ACM Prize in Computing |
| 2012 | ACM SIGOPS Mark Weiser Award (jointly with Sanjay Ghemawat) |
| 2016 | Member, American Academy of Arts and Sciences |
| 2021 | IEEE John von Neumann Medal |
| 2025 | TIME 100 Most Influential People in AI |
| 2025 | ACM SIGMOD Systems Award (as part of the Spanner team) |
The ACM-Infosys Foundation Award citation credited Dean and Ghemawat with "redefining the way in which large-scale computational systems are designed and deployed," naming MapReduce, Bigtable, and the broader Google File System and Borg infrastructure as their principal contributions [^3]. The IEEE John von Neumann Medal citation in 2021 recognized Dean's "contributions to the science and engineering of large-scale distributed computer systems and artificial intelligence systems" [^47]. The 2025 ACM SIGMOD Systems Award for Spanner recognized over thirty named contributors, including Dean, for "reimagining relational data management at global scale" [^16][^17]. Dean has also delivered keynote and invited talks at major venues including SOSP, OSDI, ISCA, NeurIPS, ICML, and ACM SIGMOD.
A substantial fraction of Dean's body of work predates the modern wave of deep learning and concerns the systems infrastructure that allowed Google to scale. Among the contributions for which he is regularly credited:
These systems collectively reflect a body of practice that has been described as the Google distributed systems school, characterized by a preference for simple data models, explicit failure handling, careful attention to constants and not just asymptotics, and the willingness to rebuild software when the next order of magnitude required it.
Dean is the subject of an internet meme in the style of Chuck Norris facts, sometimes called "Jeff Dean facts." The meme originated as an inside joke at Google during a 2007 April Fool's Day event, where co-workers wrote a list of tongue-in-cheek superlatives about him, including lines such as "Compilers don't warn Jeff Dean. Jeff Dean warns compilers," and "Jeff Dean's PIN is the last 4 digits of pi" [^48]. The list circulated externally and has been republished on engineering blogs and in technology press over the years. Dean himself has spoken about it with some embarrassment, noting that the original list was written without his knowledge and that the engineers behind most of the systems credited to him also include hundreds of colleagues. The persistence of the meme nevertheless reflects his unusual visibility within the systems and machine learning communities.
He has spoken about his career in interviews on the Lex Fridman podcast (February 2024, jointly with Noam Shazeer), in Google DeepMind's own podcast "Decoding Google Gemini with Jeff Dean" (2024), with IEEE Spectrum, and with the Google Research blog, and has given commencement and named lectures at the University of Washington, Stanford, CMU, ETH Zürich, and the Indian Institute of Science.
Dean is married to Heidi Hopper, whom he met during freshman year at the University of Minnesota; she graduated from the same university in psychology in 1990 and later earned a doctorate in organizational behavior from the University of Washington. The couple has two daughters and lives in the San Francisco Bay Area [^49][^50]. Hopper serves as president of the board of the Hopper-Dean Foundation, the family's philanthropic vehicle, which began grantmaking in 2011 and focuses on global health, education, and STEM diversity [^49]. In 2016 the foundation made gifts of approximately two million dollars each to UC Berkeley, MIT, the University of Washington, Stanford, and Carnegie Mellon to support diversity-in-STEM programs, and it has continued to support organizations including the International Rescue Committee and educational programs at the University of Minnesota.
In June 2025 Dean joined the board of the Laude Institute, a non-profit AI research organization launched by Databricks and Perplexity co-founder Andy Konwinski with an initial commitment of one hundred million dollars; the founding board also includes Turing Award laureate David Patterson and former Meta AI head Joelle Pineau [^51].
Dean's career has been the subject of profile pieces in The New York Times (2017), The New Yorker (2018, by James Somers, titled "The Friendship That Made Google Huge" and co-profiling Dean and Sanjay Ghemawat), Wired, IEEE Spectrum, and the Communications of the ACM [^11][^52]. The pieces emphasize the unusual longevity of his collaboration with Ghemawat, the breadth of systems they built, and the continuity of method between the early Google infrastructure work and the more recent AI infrastructure such as Pathways and Gemini.
The systems literature has recognized Dean's papers as foundational. The 2004 MapReduce paper, the 2006 Bigtable paper, and the 2012 Spanner paper are core readings in graduate operating systems and databases courses worldwide. The 2017 TPU paper and the 2022 Pathways paper occupy a similar place in modern computer architecture and ML systems syllabi.
Within AI research, Dean has co-authored or led papers cited tens of thousands of times. According to Google Scholar his publications have accumulated several hundred thousand citations, with an h-index well above ninety as of 2025. Beyond raw counts, his work is unusual in spanning compilers, distributed storage, distributed computation, machine learning frameworks, hardware design, and large model training, with substantive technical contributions in each.
Dean has been an outspoken advocate for energy-efficient computing and for the integration of AI methods with the natural sciences. His public talks since 2023 have argued that further substantial gains in AI capability can come from better algorithms, sparsity, hardware co-design, and improved data, and not solely from scaling parameters or compute. He was named to TIME's 100 Most Influential People in AI list for 2025 [^53].