PeopleofSuperintelligence

Researcher

Jonathan Hayase

Also known as Jon Hayase

Jonathan Hayase is a computer scientist who studies security and privacy in machine learning, tokenizers and model composition. Advised by Sewoong Oh at the University of Washington, he wrote on backdoor defenses and on merging neural networks. He was a co-author of Stealing Part of a Production Language Model, which won an ICML 2024 Best Paper award for work he did during a Google internship. His tokenizer papers include a 2024 attack that reads a language model's training mix from its BPE merge rules and the 2025 SuperBPE tokenizer. He completed his PhD in 2026, and his positions are assistant professor at UT Austin and Anthropic.

Badges

Timeline

  1. 2021

    SPECTRE backdoor defense

    In April 2021, he is first author of SPECTRE, written with Weihao Kong, Raghav Somani and Sewoong Oh. The defense uses robust covariance estimation to amplify the spectral signature of poisoned training data and remove a backdoor from the trained model.

  2. 2022

    An NSF graduate fellowship

    He wins a 2022 NSF Graduate Research Fellowship while a PhD student at the Paul G. Allen School of Computer Science & Engineering at the University of Washington, advised by Sewoong Oh.

  3. 2022

    Git Re-Basin model merging

    In September 2022, he coauthors Git Re-Basin with Samuel K. Ainsworth and Siddhartha Srinivasa. The paper introduces three algorithms that permute the units of one neural network to align it with a reference model so the two can be merged in weight space.

  4. 2023

    DataComp and training-data extraction

    In April 2023, he is among the authors of DataComp, a benchmark built on a pool of 12.8 billion image-text pairs from Common Crawl. In November 2023, he coauthors a paper showing that an adversary can extract gigabytes of training data from language models including ChatGPT.

  5. 2024

    An ICML 2024 Best Paper

    He is an author of Stealing Part of a Production Language Model, an attack that recovers the embedding projection layer of black-box models and extracts the entire projection matrix of OpenAI's Ada and Babbage for under $20. The paper receives a Best Paper award at ICML 2024, and he contributes to it during a Google internship.

  6. 2024

    Data Mixture Inference attack

    In July 2024, he is first author of Data Mixture Inference with Alisa Liu, Yejin Choi, Sewoong Oh and Noah A. Smith. The paper shows that the ordered merge rules of a BPE tokenizer reveal how much of its training data falls into each language, programming language or data source.

  7. 2025

    The SuperBPE tokenizer

    In March 2025, he coauthors SuperBPE with Alisa Liu, Valentin Hofmann, Sewoong Oh, Noah A. Smith and Yejin Choi. The tokenizer learns subwords first and then superwords that bridge whitespace, and it encodes text with up to 33% fewer tokens than BPE at a 200k vocabulary.

  8. As of October 2026

    PhD, UT Austin and Anthropic

    He completes his University of Washington PhD in 2026, with research on security and privacy in machine learning, tokenizers and model composition. His positions are assistant professor at UT Austin and Anthropic.

Career

  1. NowUniversity of Texas at AustinAssistant Professor
  2. NowAnthropic
  3. 2024GoogleIntern

Education

  1. 2026University of WashingtonPhD

From the biography.

  1. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Jonathan Hayase. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  2. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Jonathan Hayase. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.
  3. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes Jonathan Hayase. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  4. 22 of 133 people shaping AI with a former OpenAI role now work at AnthropicIncludes Anthropic. Our estimates of where OpenAI recruits came from and where 133 people with former OpenAI roles work now, including the companies and labs they founded.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

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

  1. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics arxiv.org. Checked 21 October 2026.
  2. Sewoong Oh, Paul G. Allen School of Computer Science & Engineering, University of Washington homes.cs.washington.edu. Checked 21 October 2026.
  3. About - Jonathan Hayase jon.jon.ke. Checked 21 October 2026.
  4. Git Re-Basin: Merging Models modulo Permutation Symmetries arxiv.org. Checked 21 October 2026.
  5. DataComp: In search of the next generation of multimodal datasets arxiv.org. Checked 21 October 2026.
  6. Scalable Extraction of Training Data from (Production) Language Models arxiv.org. Checked 21 October 2026.
  7. ICML 2024 Awards icml.cc. Checked 21 October 2026.
  8. Stealing Part of a Production Language Model arxiv.org. Checked 21 October 2026.
  9. Data Mixture Inference: What do BPE Tokenizers Reveal about their Training Data? arxiv.org. Checked 21 October 2026.
  10. SuperBPE: Space Travel for Language Models arxiv.org. Checked 21 October 2026.