PeopleofSuperintelligence

Researcher

Dan Valentine

Dan Valentine is an AI researcher whose work spans transformer interpretability, language-model debate and vision-language model security. In 2023, he joined MATS's Summer cohort and coauthored research on maze datasets and transformer representations. His 2024 debate paper tested whether weaker models could assess stronger ones through debate and earned an ICML Best Paper award. A 2025 ICLR study he coauthored tested more than 40 open-parameter vision-language models and found transferable image jailbreaks difficult to obtain. As of October 2026, he is a Member of Technical Staff at Anthropic, where he works on AI safety and alignment research.

Now at
Anthropic

Timeline

  1. 2023

    MATS Summer cohort

    In 2023, he joins MATS's Summer cohort, which supports 60 fellows with 15 mentors across 12 research areas.

  2. September 2023

    A maze-dataset library

    In September 2023, he coauthors “A Configurable Library for Generating and Manipulating Maze Datasets,” introducing maze-dataset for generating, processing and visualizing maze-solving datasets. The library supports rasterized and text-based outputs for convolutional neural networks and autoregressive transformers.

  3. December 2023

    Maze-solving transformers

    In December 2023, he coauthors “Structured World Representations in Maze-Solving Transformers,” which studies how small transformers represent maze topology and valid paths. The paper shows that a single token's residual stream can be linearly decoded to reconstruct an entire maze.

  4. February 2024

    Debate among language models

    In February 2024, he coauthors “Debating with More Persuasive LLMs Leads to More Truthful Answers,” which tests whether weaker models can assess stronger models through debate. The paper reports 76% accuracy for non-expert models and 88% for people. Naive baselines reach 48% and 60%, respectively.

  5. 2024

    An ICML Best Paper

    In 2024, he shares the ICML Best Paper award as a coauthor of “Debating with More Persuasive LLMs Leads to More Truthful Answers.”

  6. 2025

    Image jailbreak transfer

    In 2025, he coauthors an ICLR study of gradient-based image jailbreak transfer across vision-language models. The team tests more than 40 open-parameter VLMs and finds that transferable image jailbreaks are extremely difficult to obtain.

  7. As of October 2026

    AI safety at Anthropic

    As of October 2026, he is a Member of Technical Staff at Anthropic, where he works on AI safety and alignment research, including scalable oversight and how AI models interact with data and prompts.

  1. 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.
  2. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Anthropic. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  3. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Anthropic. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  4. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Anthropic. 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.

Sources and credits

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

  1. Dan Valentine - MATS Program Alumni matsprogram.org. Checked 22 October 2026.
  2. A Configurable Library for Generating and Manipulating Maze Datasets arxiv.org. Checked 22 October 2026.
  3. Structured World Representations in Maze-Solving Transformers arxiv.org. Checked 22 October 2026.
  4. Debating with More Persuasive LLMs Leads to More Truthful Answers arxiv.org. Checked 22 October 2026.
  5. ICML 2024 Awards icml.cc. Checked 22 October 2026.
  6. Failures to Find Transferable Image Jailbreaks Between Vision-Language Models proceedings.iclr.cc. Checked 22 October 2026.