Soumith Chintala

ref · May 25, 2026, 7:53am

Co-creator and long-time lead of PyTorch, the open-source framework powering over 90% of the world’s AI research — a Hyderabad-born engineer who was rejected by twelve US universities for graduate school and went on to build the software layer on which most of modern AI runs.

Profile

Field	Detail
Nationality	Indian-American
Born	Hyderabad, India
Current Role	CTO, Thinking Machines Lab; Affiliated Researcher, NYU
Research Areas	AI Infrastructure, Deep Learning Frameworks, Open-Source Ecosystems, Generative Models, Home Robotics
Education	B.Tech, Information Technology, VIT Vellore (2009); M.S., Computer Science, NYU (~2012)
Personal Website	soumith.ch
Blog	soumith.ch/blog
X / Twitter	@soumithchintala
GitHub	@soumith
Google Scholar	scholar.google.com

Overview

Soumith Chintala is an Indian-American AI engineer, researcher, and community builder best known as the co-creator and long-time lead of PyTorch — the open-source deep learning framework he built at Meta’s FAIR starting in 2016 that grew to power over 90% of AI research and production systems worldwide. He spent eleven years at Meta, departing in November 2025 after leaving what he described as “one of the AI industry’s most leveraged seats.” He then joined Thinking Machines Lab, Mira Murati’s AI startup, where he was appointed CTO in January 2026. His research contributions include foundational GAN papers (LAPGAN, DCGAN, Wasserstein GAN), community infrastructure work (Torch-7, EBLearn, convnet-benchmarks), and ongoing home robotics research at NYU. His story is widely cited as an illustration of persistence: he applied to twelve US universities for his master’s degree and was rejected by all before eventually gaining admission to NYU, where he studied under Yann LeCun.

Early Life & Education

Chintala grew up in Hyderabad, India, attending Hyderabad Public School. By his own account, he struggled with mathematics in his early years and came to computers and problem-solving through self-directed curiosity rather than academic excellence. He has described his path as shaped by a deep attraction to open-source software — an orientation he traces partly to growing up without access to answers, and finding that open knowledge “levels the playing field.”

B.Tech, Information Technology — Vellore Institute of Technology (VIT), Tamil Nadu, India, 2005–2009
Chintala completed his undergraduate degree at VIT Vellore, a mid-tier Indian engineering college by the standards of competitive Indian tech recruitment. After graduating, he applied to twelve US universities for a master’s degree in computer science — and was rejected by all of them. Rather than return to India or abandon the plan, he moved to the United States on a J-1 visa with no clear path forward, and eventually secured admission to NYU.

M.S., Computer Science — New York University (~2010–2012)
At NYU, Chintala was supervised by Yann LeCun, then establishing NYU’s Courant Institute as a center of deep learning research. His graduate work spanned robotics, computer vision, and early generative modeling, and included contributions to EBLearn — a C++ deep learning framework he maintained alongside LeCun and Pierre Sermanet. This period was formative: working directly under LeCun in the early years of the deep learning revival, and contributing to open-source ML tools, established both the technical depth and the community orientation that would define his subsequent career.

Career

EBLearn and Early Open-Source Work (pre-2012)

Before and during his NYU master’s, Chintala co-maintained EBLearn, a C++ framework for deep learning developed in LeCun’s lab. At its peak, EBLearn was one of the primary tools for deep learning research, and Chintala’s work on it — together with the robotics and computer vision research he conducted at NYU — constituted his entry into the open-source ML ecosystem.

Meta / Facebook AI Research (FAIR) (2014–2025)

Chintala joined Facebook AI Research in 2014, when FAIR was a small team of roughly a dozen researchers. He has described the early FAIR years — 2015 and 2016 in particular — as “probably the most productive and professionally enjoyable years of my life,” a period of working on GANs, object detection, Starcraft bots, and core infrastructure alongside Rob Fergus, Leon Bottou, Yann LeCun, Alec Radford, and others. He later rose to Vice President before departing in November 2025.

Torch-7 and convnet-benchmarks (2014–2016)
Before PyTorch, the dominant open-source deep learning framework used by Facebook, Google DeepMind, and Twitter was Torch-7 — a Lua-based framework that Chintala maintained. He also created convnet-benchmarks, a neural network benchmarking suite that became the gold standard used by NVIDIA, AMD, and Intel Nervana from 2015 to 2017 to optimize their deep learning hardware. These less-celebrated contributions established the practice of rigorous, comparative, open evaluation of ML systems that later shaped PyTorch’s design philosophy.

LAPGAN and DCGAN (2015)
In collaboration with Remi Denton, Arthur Szlam, Rob Fergus, and others, Chintala co-authored LAPGAN (NIPS 2015) — one of the first papers to generate plausible photographic images using a GAN-based approach, using a Laplacian pyramid architecture. The companion DCGAN paper (arXiv 2015, with Radford and Metz) introduced the deep convolutional GAN architecture with training stabilization techniques that made GANs practically usable for image generation. DCGAN became one of the most-cited GAN papers and effectively launched the generative image modeling era that led to Stable Diffusion, DALL-E, and subsequent image generation systems.

Wasserstein GAN (2017)
Co-authored with Martin Arjovsky and Leon Bottou, Wasserstein GAN introduced the Wasserstein distance as a training objective for GANs, substantially improving training stability and theoretical grounding. Chintala subsequently described giving up on GANs “after failing to make them stable training algorithms” — an honest assessment that also reflects his engineering ethos: if a technique cannot be reliably deployed, it does not yet deserve his full investment.

PyTorch — Co-Creation and Leadership (2016–2025)
PyTorch was created at FAIR in 2016 and released publicly in early 2017. The founding contributors included Chintala, Adam Paszke, Sam Gross, Greg Chanan, Alban Desmaison, Edward Yang, and others. The key design philosophy — dynamic computation graphs (define-by-run) rather than the static graphs of TensorFlow — made PyTorch dramatically easier to debug and iterate with, and the Python-first API lowered the barrier to entry for researchers without systems programming expertise.

Chintala served as the de facto lead of PyTorch for nearly eight years, functioning simultaneously as product manager, technical lead, communications director, community manager, and release engineer. He answered thousands of questions on the PyTorch forum and Torch forums; he has described this grassroots community investment as central to how good products get built. Under his stewardship PyTorch grew from a research prototype to the framework powering essentially every frontier AI company — including the training of GPT, Claude, Gemini, LLaMA, and Stable Diffusion — as well as production deployments at Tesla, Instagram, TikTok, Snapchat, Pinterest, and major pharmaceutical research programs.

His departure blog post (November 6, 2025) described PyTorch as handling “exascale training” and powering “foundation models that are redefining intelligence,” and confirmed the framework had achieved “90%+ adoption in AI.” He left with the explicit goal of “doing something small again” — escaping the leverage of his role to recover the feeling of building from scratch.

NYU — Affiliated Robotics Research (2019–present)

Alongside his Meta role, Chintala has collaborated with Lerrel Pinto at NYU on home robotics projects aimed at building robots that can perform household tasks without exhaustive task-specific training:

Robot Utility Models — demonstrated ~90% zero-shot accuracy on basic tasks (door opening, drawer opening, object reorientation) in unseen new environments.
On Bringing Robots Home (DobbE) — 109 tasks, tested in 10 NYC homes, 81% success rate, with 20 minutes required to learn a new task.
Additional work on tactile representations, dexterous manipulation, and semantic memory systems for home robots.

Thinking Machines Lab — CTO (November 2025–present)

After leaving Meta in November 2025, Chintala joined Mira Murati’s Thinking Machines Lab, announcing simply: “thinking machines…the people are incredible.” In January 2026, following the departure of Barret Zoph (who returned to OpenAI), Murati appointed Chintala as CTO. Murati described him as “a brilliant and seasoned leader” with over a decade of foundational contributions to AI infrastructure. His current focus is on AI infrastructure, AI research, and robotics at Thinking Machines.

Key Contributions

PyTorch — Co-created and led for nearly eight years; the open-source deep learning framework now used by 90%+ of AI researchers worldwide and powering virtually every frontier AI system in production. The dynamic computation graph design and Python-first API that Chintala championed established the dominant paradigm for ML framework design.
DCGAN (arXiv 2015, with Radford and Metz) — Introduced the deep convolutional GAN architecture with training stabilization techniques (batch normalization, specific activation choices) that made GANs practically useful; one of the most-cited generative modeling papers and the foundation for the subsequent decade of image generation research.
LAPGAN (NIPS 2015) — One of the first demonstrations of photorealistic image synthesis using GANs, using a Laplacian pyramid approach; helped establish GAN-based generative modeling as a viable research direction.
Wasserstein GAN (ICML 2017) — Introduced the Wasserstein distance training objective for GANs, substantially improving training stability and providing theoretical grounding; foundational to subsequent GAN training methodology.
Torch-7 — Maintained the pre-PyTorch Lua-based deep learning framework used by Google DeepMind, Twitter, and Meta; the direct predecessor to PyTorch and the community platform on which PyTorch’s design ideas were first tested.
convnet-benchmarks — Created the gold-standard benchmarking suite for deep learning (2015–2017) used by NVIDIA, AMD, and Intel Nervana to optimize their hardware; an early and influential example of open, comparative ML systems evaluation.
EBLearn — Co-maintained the C++ deep learning framework with Yann LeCun and Pierre Sermanet before 2012; one of the precursors to the modern ML framework ecosystem.
Home Robotics (with Lerrel Pinto, NYU) — Demonstrated zero-shot household task performance (Robot Utility Models) and scalable learning from demonstration in real homes (DobbE) — contributed to the evidence base for deployable household robots.

Awards & Recognition

TIME100 AI — Included in Time magazine’s list of the most influential people in AI.
Widely credited with enabling the modern AI era through PyTorch’s role in training GPT, Claude, LLaMA, Stable Diffusion, and most other frontier models.
PyTorch Community Impact — Chintala’s 8-year stewardship of PyTorch is widely recognized as among the most consequential individual contributions to AI infrastructure; described by colleagues as “the software layer that powers the entire AI industry.”

Key Relationships

Yann LeCun — Graduate advisor at NYU; employed Chintala at FAIR; credited by Chintala alongside Rob Fergus for “building the magical early FAIR.” LeCun’s early support of a student from a non-prestigious institution, and his sustained collaboration with Chintala at FAIR, was decisive in Chintala’s trajectory.
Adam Paszke — Core PyTorch co-creator; at the time an undergraduate in Poland who contributed to the PyTorch codebase and was invited to intern at FAIR — an example of Chintala’s openness to contributors regardless of credentials.
Alec Radford — FAIR colleague and DCGAN co-author; the GAN research period at FAIR produced some of both researchers’ most-cited early work.
Martin Arjovsky and Leon Bottou — Co-authors of Wasserstein GAN; Bottou is a senior FAIR researcher and one of the field’s mathematical forefathers.
Mira Murati — Current employer and Thinking Machines Lab CEO; appointed Chintala as CTO in January 2026.
Lerrel Pinto — NYU collaborator on home robotics; their joint work represents Chintala’s most sustained research focus outside ML infrastructure.
Mark Zuckerberg and Mike Schroepfer (Schrep) — Chintala’s departure post explicitly credited both for “believing that open-sourcing is fundamentally important and is a sound business strategy” — framing open source as a shared conviction rather than a top-down policy.

Personal Style

Chintala’s professional values can be read directly from his personal website’s list of principles: he “embraces laziness” (automate everything unwanted), “embraces simplicity,” “avoids toy or hypothetical problems,” and “rabidly loves open-source.” The last point is the most biographical: he attributes his commitment to open source to growing up without ready access to knowledge and seeing open publication of tools as a way to “equal the playing field.” His community management style — answering thousands of forum questions, staying reachable, welcoming contributions from undergraduates in Poland and grad students at NeurIPS — reflects the same egalitarian instinct that made PyTorch’s community thrive. His description of leaving Meta captures a tension that runs through his entire career: the pull between institutional leverage (leading the framework that runs all of AI) and the creative freedom of building something small and uncomfortable. His decision to leave in favor of curiosity, rather than staying for power, is consistent with the values he has articulated throughout.