Soumith Chintala credited with infrastructure

Vatsal Bhakodia said in an X post on May 23 that Soumith Chintala “built the rails” for much of today’s AI industry by helping create PyTorch, the open-source machine-learning framework now embedded across research labs and commercial model builders. The post named ChatGPT, Claude and open-source large language models as downstream beneficiaries of that infrastructure, and pointed to Chintala’s roots in Hyderabad, India. The claim is directionally supported by official material from PyTorch, OpenAI and Chintala himself, which together show how PyTorch became a common software layer for modern AI development. ### Why did one social post about Soumith Chintala travel so far? Soumith Chintala is not the public face of ChatGPT or Claude, but his work sits earlier in the stack. Chintala’s personal site says he “co-founded PyTorch,” maintained Torch-7 and spent 11 years at Meta before moving to Thinking Machines and NYU. On that same site, he writes that PyTorch “now powers most of the world’s AI research and product.” (openai.com) PyTorch’s own foundation history, in a 2022 post written by Chintala, says development began in 2016 and that the project had grown to more than 2,400 contributors and nearly 154,000 projects built on it. That post says PyTorch had become “one of the primary platforms for AI research” as well as commercial production use. ### What exactly is PyTorch in this story? (soumith.ch) PyTorch is the software framework researchers and engineers use to build, train and deploy neural networks. Meta’s open-source page describes it as an open-source machine-learning framework that accelerates the path “from research prototyping to production deployment.” Meta’s 2018 engineering post introducing PyTorch 1.0 described the same goal: combining research flexibility with production-oriented capabilities. (pytorch.org) That matters because the companies making headline AI products often depend on frameworks that are less visible than the products themselves. The social post praised Chintala for infrastructure work, and the official record supports that framing more than any claim that he directly built ChatGPT or Claude. ### Is there direct evidence tying PyTorch to ChatGPT and OpenAI models? (opensource.fb.com) OpenAI gave the clearest direct confirmation. In a January 30, 2020 post, OpenAI said, “We are standardizing OpenAI’s deep learning framework on PyTorch,” adding that the shift made it easier for teams to create and share optimized implementations of models. OpenAI said PyTorch had cut iteration time on some generative-modeling research “from weeks to days.” (opensource.fb.com) OpenAI’s 2025 model card for its open-weight gpt-oss models goes further on training details, saying those models were trained on NVIDIA H100 GPUs using the PyTorch framework with Triton kernels. That does not by itself document every ChatGPT model, but it does show PyTorch inside OpenAI’s model-training stack. (openai.com) ### What about Claude and the broader open-source model world? Anthropic’s public engineering and careers pages describe the company as building large AI systems and list substantial research, compute and infrastructure hiring, but the pages surfaced here do not explicitly state “Claude is trained in PyTorch.” That means the strongest verified part of the social claim is OpenAI’s documented PyTorch use and PyTorch’s broad industry role, not a fresh Anthropic admission in the material reviewed. (cdn.openai.com) The broader open-source case is easier to document. PyTorch’s foundation post says companies including AMD, AWS, Google Cloud, Hugging Face, Microsoft Azure and Nvidia have invested in the project, while the current PyTorch site describes a large ecosystem of community tools and maintainers. ### Why was Bhakodia’s phrasing about “infrastructure” the key point? (anthropic.com) Infrastructure is the most defensible word in the post. Chintala’s work was not the chatbot interface, the product launch or the consumer brand; it was the framework layer that many labs adopted to move from experimentation to training and deployment. Meta’s documentation and OpenAI’s own statements both describe PyTorch in exactly those terms. (pytorch.org) PyTorch’s next public milestone is PyTorch Conference North America on October 20-21, 2026, in San Jose, according to the project’s homepage. Chintala’s biography page on that site remains live, and the PyTorch Foundation continues to operate under the Linux Foundation structure announced in 2022. (pytorch.org) (opensource.fb.com)