PeopleofSuperintelligence

Builder

James Bradbury

James Bradbury's career links deep-learning software with research on sequence models and large language models. At Salesforce Research, he worked on natural-language-processing models and coauthored Quasi-Recurrent Neural Networks. He coauthored the JAX software release, the PyTorch paper and PaLM, and contributed ideas to How to Scale Your Model about the systems view of the Transformer. As of October 2026, he is Head of Compute at Anthropic, where almost one million Trainium2 chips train and serve Claude.

Now at
Anthropic
Works
16

Publication numbers from OpenAlex, as of 15 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

  1. 2016

    Salesforce Research and QRNNs

    In April 2016, Salesforce acquires MetaMind, and he joins Salesforce Research. There, he works on deep-learning models for natural-language processing. He coauthors Quasi-Recurrent Neural Networks, which pair parallel convolutional layers with recurrent pooling for sequence modeling.

  2. By 2017

    A Stanford linguistics degree

    By 2017, he holds a degree in linguistics from Stanford University.

  3. 2018

    JAX software release

    In 2018, he coauthors the JAX software release, titled “JAX: composable transformations of Python+NumPy programs.”

  4. 2019

    The PyTorch paper

    In 2019, he coauthors the NeurIPS paper on PyTorch, a machine-learning library with an imperative, Pythonic style and GPU support.

  5. 2022

    PaLM and Pathways

    In 2022, he coauthors PaLM, a 540-billion-parameter language model trained on 6,144 TPU v4 chips with Pathways.

  6. Since 2023

    Joining Anthropic

    He joins Anthropic as Head of Compute in 2023 from Google DeepMind's JAX Engagements Team, where he works with large-scale LLM users.

  7. 2025

    How To Scale Your Model

    In 2025, he and Reiner Pope, Noam Shazeer, and Blake Hechtman originate many ideas in the book and show an early understanding of the systems view of the Transformer.

  8. As of October 2026

    Head of Compute at Anthropic

    As of October 2026, he is Head of Compute at Anthropic, where almost one million Trainium2 chips train and serve Claude.

Sources and credits

Photo: We have no published photo of James Bradbury yet. Send us a free-licensed photo.

  1. Author: James Bradbury | NVIDIA Technical Blog developer.nvidia.com. Checked 14 October 2026.
  2. Quasi-Recurrent Neural Networks arxiv.org. Checked 14 October 2026.
  3. JAX: composable transformations of Python+NumPy programs bibbase.org. Checked 14 October 2026.
  4. PyTorch: An Imperative Style, High-Performance Deep Learning Library proceedings.neurips.cc. Checked 14 October 2026.
  5. PaLM: Scaling Language Modeling with Pathways research.google. Checked 14 October 2026.
  6. Keynote Speakers | PyTorch Conference events.linuxfoundation.org. Checked 14 October 2026.
  7. Conclusions and Further Reading | How To Scale Your Model jax-ml.github.io. Checked 14 October 2026.
  8. AI Accelerator - AWS Trainium Customers - AWS aws.amazon.com. Checked 14 October 2026.