Created PyTorch, Meta AI research
Soumith Chintala
Profile
Soumith Chintala co-created PyTorch — the framework that runs most of AI research and a growing share of production AI. If you’ve ever written import torch, you’ve used his work. PyTorch wasn’t the obvious winner when it shipped in 2016. TensorFlow had Google’s weight behind it and was the safe enterprise bet. PyTorch won anyway because it felt natural to write, debugged like normal Python, and trusted researchers to know what they wanted. That bet — usability over corporate polish — shaped a decade of AI.
He spent eleven years at Meta (then Facebook AI Research), rising to VP of AI Infrastructure and Meta Fellow. Along the way he co-authored DCGAN with Alec Radford and Luke Metz — one of the foundational GAN papers that made generative models actually trainable — and later Wasserstein GAN, which fixed a lot of what DCGAN couldn’t. He’s been publicly honest about eventually giving up on GANs as a stable training paradigm — a rare thing in a field where people rarely admit when a line of work doesn’t pan out.
In November 2025 he left Meta. In January 2026 he joined Thinking Machines Lab, Mira Murati’s startup, as CTO. The move signals something: the center of gravity in open AI infrastructure is shifting from the hyperscalers to a new crop of labs staking out ground between frontier closed models and the research commons.
For developers learning AI, Soumith is worth following because he thinks out loud about the full stack — the framework, the hardware, the community, the trade-offs. He grew up in Hyderabad, went to a tier-2 engineering school (VIT), got rejected from most of his grad school applications, did his MS at NYU, and then spent a decade quietly building the tool everyone else builds on top of. His story is a useful counter to the Stanford-pedigree narrative of how AI gets made.
Key Articles & Papers
PyTorch: An Imperative Style, High-Performance Deep Learning Library Unsupervised Representation Learning with Deep Convolutional GANs (DCGAN) Wasserstein GAN Automatic Differentiation in PyTorch Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks (LAPGAN) Leaving Meta and PyTorch Open Source AI is AI We Can Trust — Latent Space interview Soumith Chintala on The Gradient Podcast Soumith's personal siteSpotify Podcasts
