Meet Soumith Chintala

Summary of Soumith Chintala and his portfolio site

Soumith Chintala is an AI researcher, engineer, open-source community builder, and occasional investor. His homepage presents him as based in New York, originally from Hyderabad, India, and educated at NYU and VIT Vellore. He says he is currently at Thinking Machines and NYU, working on AI infrastructure, AI research, and robotics, after spending 11 years at Meta.

A central part of Soumith’s work is AI tooling and open source. He co-founded PyTorch, maintained Torch-7, and has spent a large part of his career building developer tools and communities for machine learning researchers and engineers. In his post “Leaving Meta and PyTorch”, he reflects on leading PyTorch from zero to “90%+ adoption in AI,” describing it as infrastructure now used for exascale training, foundation models, production systems, and AI education around the world.

His blog also gives a useful window into his product and engineering philosophy. In “PyTorch’s design origins”, he traces PyTorch’s roots through Torch7, Chainer, and HIPS/autograd, emphasizing careful attribution and the evolution of ideas across open-source systems. In “Growing open-source from Torch to PyTorch”, he explains his approach to growing open-source projects: stay stubborn about core principles, focus on a specific user group first, avoid misleading metrics, and build strong feedback loops with users. He also writes about large-scale AI systems, such as in “How to train a model on 10k H100 GPUs?”, which summarizes practical concerns in frontier-scale training: parallelism, communication, checkpointing, hardware failures, and recovery.

Beyond infrastructure, Soumith is strongly interested in home robotics. His homepage says he wants to build household robots that can help with chores, and he frames this as requiring world simulation, multimodal models, generative models, and efficient representations of human-centric environments. He highlights recent robotics work with NYU collaborator Lerrel Pinto, including:

• Robot Utility Models — zero-shot robotic utility models for tasks like door opening, drawer opening, and object reorientation.
• On Bringing Robots Home / Dobb-E — home robot learning across 109 tasks in 10 NYC homes.
• Dexterity from Touch — self-supervised tactile representation learning through robotic play.
• CLIP-Fields — semantic fields for robotic memory.
• Holo-Dex — teaching dexterity using immersive mixed reality.

The site is compact and personal: a homepage, a blog, and a media page. The media page collects podcasts, talks, and press coverage around PyTorch, open-source AI, AI infrastructure, robotics, GANs, and his move from Meta to Thinking Machines. It shows that he is active not only as a builder and researcher, but also as a public communicator about AI systems and open-source ecosystems.

He also notes that he invests mainly within his network, including companies such as Runway, 1X, Osmo, Anthropic, Together.ai, and Lepton.

Relevant profiles linked from the site include GitHub, X, LinkedIn, and Google Scholar / Publications.