Soumith Chintala

Soumith Chintala is an AI researcher and engineer who co-founded PyTorch, maintained Torch-7, and co-authored research on generative adversarial networks and household robotics. In March 2024, he was Meta’s engineering lead for AI infrastructure, PyTorch, and generative AI. His personal website, as of September 18, 2026, lists affiliations with Thinking Machines and NYU after eleven years at Meta, with a focus on AI infrastructure, research, and robotics.[2][1]

Chintala describes his approach as incremental ambition: begin with something simple, then add complexity while keeping common workflows on an easy path. Advanced uses remain possible but may require extra steps. In his March 2024 Latent Space interview, he explained PyTorch’s complexity through the problem of fitting computations onto hardware: the best implementation changes with tensor shapes, operations, and memory hierarchies. Many specialized operator configurations help deliver throughput without lengthy compilation waits. In his assessment, a framework covering a narrower family of workloads can be much simpler.[1][2]

He described hardware support as a response to user demand: “We very much don't want to king make the hardware side of things.” Apple GPU support, he recalled, involved both Apple engineers and substantial work by Meta-funded engineers. He applied similar criteria to dependencies—ease of installation, smooth integration, and benefits to users—citing PyTorch’s use of OpenAI Triton. For Mojo, he saw a possible role in letting torch.compile consume custom Mojo subgraphs; using Mojo as the complete frontend would instead replace PyTorch. This was a proposed interoperability use case, not an announced integration.[1]

His infrastructure role also extended to Llama 2. Chintala credited Guillaume Lample and team for Llama 1 and said he had little involvement in that work, while helping Llama 2 with infrastructure needs. Discussing model development in March 2024, he emphasized the timing of successive generations: whether better, cleaner data made it worth starting the next model rather than continuing to train the previous one.[1]

Chintala links his commitment to open source to growing up in India with limited internet access. Decentralized knowledge, he said, helped him learn faster without paying: “I'm irrationally interested in open source.” His argument for open AI centers on broad access and participation: people need access to the technology to contribute feedback, and decisions made within a narrow cultural bubble may fail to reflect users elsewhere. In that discussion, he used “open source” broadly to mean open availability, acknowledging uncertainty about what constitutes source for model weights and differences among licenses.[1]

He also identified a coordination problem. Open models could be used through many separate applications without their users’ feedback returning to model developers. His March 2024 proposal combined feedback controls across those applications, a shared collection service, and filtering to remove spam and retain useful responses. The point was to learn from distributed, real-world use. He regarded a single comparison interface as too narrow to capture the range of uses, including coding assistance and character-based interactions.[1]

At NYU, Chintala works on home robotics with Lerrel Pinto and team. In March 2024, he described spending Wednesday afternoons there and prioritizing household user experience, few demonstrations, and reliable hardware. The team was exploring tasks learned from two or three demonstrations, with restricted generalization and necessary starting conditions. He argued that commercialization also required progress on motors, sensors, and maintenance: a robot that performs a demonstration but repeatedly needs repairs presents a different problem from a dependable household appliance.[1][2]

Chintala co-authored the 2023 Dobb·E project with Nur Muhammad “Mahi” Shafiullah, Anant Rai, Haritheja Etukuru, Yiqian Liu, Ishan Misra, and Lerrel Pinto. Its demonstration tool, “The Stick,” combined a reacher-grabber, printed parts, and an iPhone. The team pretrained visual representations on thirteen hours of interactions collected in twenty-two homes. For a new task, the workflow used five minutes of demonstrations and fifteen minutes of adaptation before deployment on a Stretch robot. The project reported 81% success across 109 tasks in ten evaluation homes; that result describes the experiment, rather than general household reliability. The team released its software, models, data, and hardware designs.[9]

Chintala’s earlier work in deep-learning frameworks, benchmarking, and image generation supplies context for these later systems and robotics contributions.[2][7][8][1]

Chintala grew up in Hyderabad and studied at VIT Vellore and NYU. Before 2012, he maintained the C++ deep-learning framework EBLearn with Pierre Sermanet and Yann LeCun; he later maintained Torch-7. His account of PyTorch’s origins, published in December 2023, places its groundwork in early 2016 among Torch7 contributors collaborating online.[2][5]

He created convnet-benchmarks to compare public convolutional-network implementations. Its ImageNet-model tests averaged forward and backward pass times over ten runs on a specified machine, omitting dropout and softmax. The repository warned that entries using cuDNN could be mistaken for comparisons of native framework kernels. Chintala later recalled startups using the results in pitch decks and credited Scott Gray’s fast convolution kernels for Nervana’s strong showing. By March 2024, his criticism of inference-provider benchmarks extended beyond speed: comparisons should include realistic workloads, cost at scale, and latency and reliability sampled across different days and times.[10][1][2]

With Alec Radford and Luke Metz, Chintala co-authored DCGAN, initially submitted in November 2015. The authors introduced constrained convolutional GAN architectures and reported learned representations ranging from object parts to scenes. With Martin Arjovsky and Léon Bottou, he co-authored Wasserstein GAN, initially submitted in January 2017. That paper proposed an alternative training algorithm and reported improved stability, reduced mode-collapse problems, and learning curves useful for debugging and hyperparameter search.[7][8]

Chintala has offered distinct accounts of leaving GAN research: his website says he gave it up after failing to make the training algorithms stable; in March 2024, he also described difficulty judging his contribution in a crowded field where others were pursuing similar ideas. His November 6, 2025 essay, “Leaving Meta and PyTorch”, expressed another boundary on his career: “I don't want to be doing PyTorch forever.”[2][1][4]