Without infighting, Meta was unable to retain the father of PyTorch

Soumith Chintala said he doesn’t want to be tied to the same project for decades. Just now, Soumith Chintala, the father of PyTorch, announced that he will officially leave Meta on November 17th, ending his 11 year career journey.

I don’t want to be PyTorch for the rest of my life, “he wrote in his farewell letter.” I want to start over and do some small and unknown things. I don’t want to carry the hypothetical regret of ‘if I never tried something outside Meta, would I regret it?’. ”

As soon as the news came out, the entire AI community instantly went viral, and a group of big shots sent their blessings:

LeCun said, “Wishing you all the best in your next job

Andrej Karpathy: “May you once again welcome your golden age in the next field that truly ignites your passion! ”

The engineer who brought PyTorch from a laboratory project to mainstream AI platforms around the world reviewed his 11 year journey at Meta in a heartfelt resignation letter and explained why he chose to say goodbye at this moment.

Here is the complete farewell letter he posted.

Leaving Meta and PyTorch

I will leave Meta on November 17th and step down as the head of PyTorch.

I don’t want to be PyTorch forever. After returning from a long vacation, I realized that the project could operate on its own without me, and it seemed like the perfect time to leave.

During my eleven years at Meta, it almost occupied my entire career. I have made many friends. For nearly eight years, I led PyTorch from scratch, and its usage rate has now exceeded 90% in the AI field.

Leaving this matter is one of the most difficult decisions in my life. But I left with a heart full of gratitude.

Nowadays, PyTorch can support exascale training, driving the redefinition of intelligent foundational models, and almost all major AI companies are using it in production environments.

PyTorch is widely used from MIT classrooms to rural schools in India. I once dreamed of making AI tools easier to use, and now it has truly come true. The threshold that I once hoped to lower has now been almost completely eliminated. Of course, there are still many things to be done. As long as AI continues to evolve at an astonishing pace, PyTorch will continue to catch up.

We often focus on the future that has not yet arrived, to the point where we forget how far we have already come. Thank you to all those who have fought alongside me, those who believe that scientific research should be fun, tools should be elegant, and open source can change everything.

This is not my journey, but our journey.

What will I do next?

Some small things, some new things, some things that I don’t fully understand, some things that make me feel uneasy. I could have changed direction within Meta, but I want to see the outside world. I want to start over and do some small and unknown things. I don’t want to carry the hypothetical regret of not regretting if I never try.

Leaving is indeed difficult. I may now be sitting in one of the most influential positions in the entire AI industry, with the software layer I am responsible for supporting the entire AI industry; Every mainstream AI company and hardware manufacturer maintains direct contact with me.

This kind of influence is difficult to let go of.

But in the end, curiosity won in my heart. Please continue to make AI delicious and accessible, I will continue to follow and perhaps occasionally raise issues. I won’t really leave, just continue to participate in a different way.

PyTorch no longer needs me

I don’t want to spend my whole life doing PyTorch. I don’t want to become like Guido (father of Python) or Linus (father of Linux), tied to the same project for decades. Last November, at the same time as my daughter was born, I began planning my departure. My goal is clear: to put PyTorch in a stable and healthy state before leaving.

By August of this year, which is the second half of my parental leave, I realized that Edward, Suo, Alban, Greg, John, Joe, and Jana were all ready. The team no longer relies on me to solve difficult personnel, product, technical, and organizational issues as they did in the past. The high-risk points that I once highlighted in red are now improving.

This project no longer requires me.

Unlike in 2020-2022 (when I temporarily left to work on a robotics project and later returned to take over due to Lin, Dima, and Dwarak’s departures), this time I have full confidence that PyTorch has truly demonstrated resilience.

The people who best represent PyTorch culture, Greg, Alban, Ed, Jason, and Joe, are now at the decision-making level. And partners with highly aligned values, Suo, John, and Jana, have also joined in. More importantly, there is a long list of people who share the same values and are willing to take over the baton in the future.

There are many details that give me confidence in this team:

John has been working in the Julia language and open source community for many years (in fact, we studied Torch. jl together as early as 2015);

Suo has been my strongest system builder and strategic partner over the past two years;

Jana has extensive experience in building high reliability core systems for a long time. Over the past few months, we have had numerous in-depth discussions at both the technical and organizational levels, which has given me great peace of mind.

In addition, the product planning and execution results in 2025 will be the best proof of PyTorch’s strength.

I have full confidence in this PyTorch team, they will do an excellent job. Perhaps PyTorch’s style will change, as I no longer influence its style from top to bottom, but I believe its core values will remain unchanged and the product will continue to excel.

My time at Meta

In the early days of FAIR (Facebook AI Research), it was truly a magical time. At that time, I was in a small family composed of extremely intelligent people, working together in a completely open environment to create cutting-edge AI technology.

I have studied GANs with Emily Denton, Rob Fergus, Leon Bottou, Martin Arjovsky, and now legendary Alec Radford; Developing StarCraft intelligent agents together with Gabriel Synnaeve; Built the first FAIR cluster with Howard Mansell; Researching object detection with Adam Lerer and Piotr Dollar; And ultimately created PyTorch.

The joy and passion of that time are hard to describe in words.

2015 and 2016 were probably the most fulfilling and joyful years of my career, and I think I will always look back on those days with romantic feelings. When I first joined FAIR, I had strong imposter syndrome and the first three months were very difficult. I would like to sincerely thank Andrew Tulloch, who is the most considerate, kind, and wise mentor. Without his help, I might not have been able to persevere. Just because he returned to Meta, I am full of confidence in Meta’s future.

Thank you everyone

I love building every part of it, whether it’s designing, managing, serving as a product manager (PM), technical lead (TL), external communication lead, documentation engineer, release engineer, fixing countless bugs, conducting growth experiments, building it into a unified product with hundreds of people collaborating, and ultimately transitioning it from a research tool to an open source project maintained by the entire industry. I am fully committed to it.

To the core PyTorch team of Meta: engineers, researchers, open source maintainers, document writers, CI infrastructure personnel, hardware partners, community builders, and hundreds of contributors both inside and outside Meta. Thank you. You turned a library into a movement.

There are too many people worth thanking and paying tribute to, but I have to specifically mention: Adam Paszke、Sam Gross、Greg Chanan、Joe Spisak、Alban Desmaison、Edward Yang、Richard Zou、Tongzhou Wang、Francisco Massa、Luca Antiga、Andreas K ö pf, Zach DeVito, Zeming Lin, Adam Lerer, Howard Mansell, Natalia Gimelstein, and Schrep. They made the first release of PyTorch a reality.

Later on, more people became important: Lu Fang、Xiaodong Wang、Junjie Bai、Nikita Shulga、Horace He、Mark Saroufim、Jason Ansel、Dmytro Dzhulgakov、Yangqing Jia、Geeta Chauhan、Will Constable、Briah Hirsh、Jane Xu、Mario Lezcano、Piotr Balecki、Yinghai Lu、Less Wright、Andrew Tulloch、Bruce Lin、Woo Kim、Helen Suk、Chris Gottbrath、Peng Wu、Joe Isaacson、Eli Uriegas、Tristan Rice、Yanan Cao、Elias Ellison、Animesh Jain、Peter Noordhuis、Tianyu Liu、Yifu Wang、Lin Qiao， And hundreds of other equally important people.

If I couldn’t write down every one here, it would be a sin.

The reason why PyTorch is great is entirely because of these people.

In the process of building PyTorch, the happiest moment for me was encountering users who enthusiastically shared their joy, love, and feedback.

I still remember in 2017, at NeurIPS, a graduate student with a voice so excited that he choked up said to me: He had been struggling for three years for his research, but within three months of using PyTorch, he had made tremendous progress and was finally ready to graduate.

At that moment, I deeply realized that what we do truly holds great significance for many people, even if they may not always express it.

I do miss the intimate atmosphere of the PyTorch community at the beginning, when the conference had only 300 people and was like a big family gathering. But I also believe that the loss of this intimacy is a small price paid for greater meaning. After all, the PyTorch conference has now expanded to a scale of 3000 people, and the cooperation reached on-site can even affect the entire industry market; More importantly, it is helping thousands of people complete their best AI jobs.

I miss that intimacy, but I am also immensely proud of this growth.

To Mark Zuckerberg and Mike Schroepfer. Thank you for always believing that open source is not only crucial, but also a reasonable and visionary business strategy. It is actually very difficult to understand this in business operations, but we have always been in sync on this path and have never needed to say much. Without you, there would be no FAIR or PyTorch. And both of these are of great significance to me.

To Yann LeCun and Rob Fergus, thank you for creating that magical early FAIR, which I still deeply admire today.

To Aparna Ramani, an extremely rare outstanding leader in Meta. You are always able to set extremely high standards for the organization, combining excellent technical insight with macro strategic thinking. You can discuss both underlying infrastructure and industry strategy in one conversation, and your execution is extremely strong! I have learned too much from you.

To Santosh, Kaushik, Delia, Oldham, and Ben, thank you for your enthusiastic acceptance of the infrastructure team. For someone from FAIR with a diverse cultural background, you have made me feel the warmth of home and truly become a member of the team. Thank you sincerely.