Soumith Chintala

Soumith Chintala grew up in Hyderabad, India, dreaming of writing code that mattered. He studied at VIT Vellore and then crossed the Atlantic to New York for his master's degree at NYU under Yann LeCun, one of the fathers of deep learning. The path wasn't easy: before landing at Meta, he collected 27 rejections from US doctoral programs. He repeats that number himself in interviews, not as a complaint, but as proof that failure is just information, not a verdict.

In 2016, together with a small team at Facebook AI Research (FAIR), he launched PyTorch, a deep learning framework designed so researchers could run experiments fast, debug with ease, and think in clean code instead of wrestling with infrastructure. The difference from what existed at the time was enormous: PyTorch treated researchers like real programmers, not users who needed a black box. Today, the vast majority of the world's AI papers are written in PyTorch. That's not hyperbole: it's what the publication statistics from NeurIPS and arXiv show. He also co-authored the seminal DCGAN paper (2015) with Alec Radford and Luke Metz, and the Wasserstein GAN paper (2017) with Martin Arjovsky and Leon Bottou, two cornerstones of image generation with AI.

He spent eleven years at Meta as one of the company's most respected engineers, reaching the title of VP Fellow of Engineering. That's a distinction Meta grants to a very small number of people: those who define the company's technical future. In 2025 he left Meta to join Thinking Machines Lab as CTO, Mira Murati's ambitious project that raised two billion dollars before launching its first product, with the promise of building AI that is genuinely useful to the world.

What strikes me as powerful about Soumith isn't just what he built, but how he made it available. PyTorch is open source because he and his team believed research moves faster when everyone can see the engine. That decision helped thousands of researchers at small universities and under-resourced startups compete with the big labs. Every time you use Claude, or any modern AI model, there's a good chance that at some point in its training, PyTorch code ran under the hood. That's a legacy.