리타🌶
2.2K posts

리타🌶
@youvngdae
Just like a moon in the day, just because they are invisible does not mean they do not exist. I will always be with you, whenever or wherever.



My only Telegram username is : t.me/webeeu ✅ Any other account is a scam. Please ignore them! Thanks!


A kid from Hyderabad, India who could not afford his own computer in college built the framework that now runs 80% of all AI research on Earth, and last month he quietly walked away from Meta to start something nobody can figure out yet. His name is @soumithchintala , and the framework is called PyTorch. Almost every model you have ever heard of runs on his code. GPT was trained on it. Claude was trained on it. Llama, Stable Diffusion, Tesla Autopilot, every Hugging Face model, every research paper at NeurIPS for the last five years. All of it sits on top of a library a graduate student wrote because the existing tools were too painful to use. Here is the story, because almost nobody outside the AI research community knows who he is. Soumith grew up in Hyderabad in the 1990s. His family did not have a computer at home. He used to walk to a neighbor's house to use theirs. He learned to code on whatever machine he could borrow. He got into @VIT_univ on merit and graduated in 2009 with a degree in electronics and a hunger for computer vision research he could not afford to pursue in India. He left for the United States that fall. @nyuniversity graduate school. Yann LeCun's lab. The same lab that would later become the seed of Meta's AI division. In 2014 he joined @facebook AI Research as one of the first employees. The field was a mess. Researchers were using a framework called Torch that ran on Lua, a programming language almost nobody outside the gaming industry used. Everyone hated it. TensorFlow was the official Google framework and it was so painful to debug that researchers were quietly switching back to writing models in raw NumPy. Soumith looked at this and decided to build something better. He pitched PyTorch internally at Facebook in 2016. Everyone told him to drop it. TensorFlow was the standard. @Google had thousands of engineers behind it. Building a competing framework was political suicide inside a company that was supposed to be cooperating with Google's research ecosystem. He shipped it anyway. January 2017. A small team. A clean Python API. Dynamic computation graphs that let researchers debug their models the same way they debug regular code. No new language. No new compiler. Just Python that worked the way Python is supposed to work. The research community switched within 18 months. By 2019, half of all papers at major AI conferences were using PyTorch. By 2022, it was 80%. By 2024, TensorFlow was essentially dead in research and barely hanging on in production. OpenAI moved to PyTorch. Tesla moved to PyTorch. Every Chinese AI lab moved to PyTorch. Anthropic was built on PyTorch from day one. The library Soumith built in his spare time inside a company that told him to stop became the operating system of modern artificial intelligence. In 2022 Meta transferred PyTorch to the Linux Foundation. @soumithchintala pushed for this. He did not want any one company to own the framework that the entire field depended on. Google never did this with TensorFlow. He gave away the most strategically important software asset Meta had ever built because he thought open science required it. In November 2024 he announced he was leaving Meta after 11 years. He has not said what he is building next. He is still on the PyTorch board. He still commits code on weekends. He has been giving talks at universities about hardware bottlenecks in AI training and the failure of current GPU memory architectures to keep up with model size. Researchers who have had dinner with him say he is obsessed with something he calls compute democratization, the idea that the next generation of AI infrastructure has to run on hardware regular people can actually afford. Whatever he ships next will probably be free. A kid who could not afford a computer built the framework every AI lab on Earth runs on, gave it away, and left the highest-paid job in tech to start over. The big labs spent a decade trying to lock down AI behind APIs and paywalls. He just keeps shipping the thing that makes their walls irrelevant











