TensorFlow had a 3-year head start. Keras was the "easy" option. Caffe had the research crowd. And PyTorch — Facebook's scrappy upstart from 2017 — ate them all.
By 2024, over 75% of new research papers on arXiv use PyTorch. OpenAI, Anthropic, Meta, Tesla Autopilot, Stability AI — all PyTorch. It's not even close anymore.
Why? One word: Pythonic. PyTorch feels like writing normal Python. TensorFlow felt like learning a new programming language inside Python. That friction killed it.
Build in public
Tweet: 'Starting my ML journey with PyTorch. The framework researchers and industry actually use.'
Rex — Why PyTorch Won
TensorFlow had a 3-year head start. Keras was the "easy" option. Caffe had the research crowd. And PyTorch — Facebook's scrappy upstart from 2017 — ate them all.
By 2024, over 75% of new research papers on arXiv use PyTorch. OpenAI, Anthropic, Meta, Tesla Autopilot, Stability AI — all PyTorch. It's not even close anymore.
Why? One word: Pythonic. PyTorch feels like writing normal Python. TensorFlow felt like learning a new programming language inside Python. That friction killed it.
Build in public
Tweet: 'Starting my ML journey with PyTorch. The framework researchers and industry actually use.'