Soumith Chintala is chief technology officer of Thinking Machines Lab and a co-founder of PyTorch, the open-source framework that transformed how researchers build, debug, and deploy machine-learning systems. His work spans deep-learning infrastructure, foundational generative-model research, household robotics, and personal AI assistants designed to preserve their users’ privacy and control.
Building the foundations
Chintala grew up in Hyderabad, studied at VIT Vellore, and earned a computer-science master’s degree at New York University. In Yann LeCun’s laboratory, he worked on pedestrian detection, optical character recognition, and depth imagery. He subsequently developed mobile music- and vision-related machine-learning systems at MuseAmi before joining Facebook AI Research in 2014.
Early contributions to the C++ framework EBLearn, the Lua-based Torch-7, and convolutional-network benchmarking exposed him to the practical difficulties researchers faced when building neural networks. At Facebook, later Meta, he helped create PyTorch, whose Python-first design, dynamic computation, and imperative execution made models easier to inspect and modify without abandoning performance. Its foundational paper, which Chintala co-authored, describes the framework’s emphasis on usability, extensibility, and acceleration.
He also co-authored influential generative-modeling research, including DCGAN, which applied convolutional adversarial architectures to visual representation learning, and Wasserstein GAN, which introduced an alternative training objective intended to improve stability and make learning progress easier to diagnose.
After eleven years at Meta and approximately eight leading PyTorch, Chintala announced his departure in November 2025. He became Thinking Machines Lab’s chief technology officer in January 2026 and maintains an affiliation with NYU.
From open infrastructure to useful autonomy
Open-source infrastructure as access. Chintala approaches developer experience, community support, and high-performance systems as interconnected engineering problems. PyTorch gave researchers and companies a shared foundation they could inspect and adapt; he expects open models to benefit from similar coordination as contributions compound across organizations.
Household robotics outside the laboratory. With NYU collaborators, Chintala co-authored Dobb·E, an open-source household-manipulation system evaluated across 109 tasks in ten homes, reporting an 81 percent success rate. The project combines inexpensive demonstration hardware with adaptation to new tasks in approximately twenty minutes. His work on Robot Utility Models pursues zero-shot household robotics, enabling robots to attempt familiar categories of manipulation in unfamiliar environments without additional fine-tuning.
Local, privacy-preserving AI agents. Chintala defines an agent by its ability to act, not simply retrieve information. Reliable personal assistants therefore need access to fragmented context across messaging, email, calendars, and financial services; missing information can make an otherwise capable system dangerously misleading. He favors an always-on Mac mini as a practical local host, avoiding wearable battery limits and mobile operating-system restrictions while keeping sensitive context under the user’s authority.
Catastrophic-action safeguards. In his AI Engineer Summit session, Chintala identifies unreliable computer use, weak visual personalization, limited open-source voice interfaces, and consequential delegated actions as unresolved obstacles. Agents must distinguish reversible mistakes from expensive purchases or other harmful decisions and request human approval when the stakes demand it.
