PeopleofSuperintelligence

Researcher

Tom Conerly

Also known as Thomas Conerly

Tom Conerly is an AI researcher whose work at Anthropic spans language model alignment and interpretability. In 2022, they coauthored studies on human-feedback alignment, repeated data, model self-knowledge, and Constitutional AI. From 2023 to 2025, they worked on sparse autoencoders, cross-layer features, and circuit tracing for Claude 3 Sonnet and Claude 3.5 Haiku. An interference-weights note they coauthored in July 2025 examined their effect on global circuit analysis. In May and July 2026, they reviewed Anthropic interpretability papers on language autoencoders and verbalizable representations.

Now at
Anthropic

Timeline

  1. April 2022–May 2022

    RLHF and repeated data

    In April 2022, they coauthor Anthropic's Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback, which uses preference modeling and RLHF to fine-tune language models as helpful and harmless assistants. In May 2022, they coauthor Scaling Laws and Interpretability of Learning from Repeated Data, which finds that repeated data can raise test loss mid-training and harm copying and induction heads.

  2. July 2022–December 2022

    Self-knowledge and Constitutional AI

    In July 2022, they coauthor Language Models (Mostly) Know What They Know, which studies whether language models can assess their claims and predict which questions they can answer correctly. In December 2022, they coauthor Constitutional AI: Harmlessness from AI Feedback, which trains a harmless assistant through self-improvement without human labels identifying harmful outputs.

  3. October 2023

    Towards Monosemanticity

    In October 2023, they coauthor Towards Monosemanticity, which uses a sparse autoencoder to extract interpretable features from a one-layer transformer.

  4. May 2024

    Scaling features in Claude 3 Sonnet

    In May 2024, they are a core contributor to Scaling Monosemanticity, which extends sparse autoencoders to Claude 3 Sonnet. They lead its Scaling Dictionary Learning section, scale up activation collection with Adly Templeton, and run scaling law experiments with Jack Lindsey and Tom Henighan.

  5. October 2024

    Sparse crosscoders

    In October 2024, they are one of four core contributors to Anthropic's Sparse Crosscoders for Cross-Layer Features and Model Diffing. The note introduces crosscoders that read and write across model layers to produce shared features.

  6. March 2025

    Circuit tracing and model biology

    In March 2025, they coauthor Circuit Tracing: Revealing Computational Graphs in Language Models and the companion On the Biology of a Large Language Model. They improve dictionary-learning efficiency and optimize the sparsity penalty and nonlinearity for cross-layer transcoders.

  7. July 2025

    Interference weights in toy models

    In July 2025, they coauthor A Toy Model of Interference Weights with Chris Olah and Nicholas L. Turner. They rewrite sections to improve communication in a note on how interference weights affect global circuit analysis.

  8. May 2026–July 2026

    Anthropic interpretability review

    In May 2026, they are an internal reviewer for Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations. In July 2026, they help review Verbalizable Representations Form a Global Workspace in Language Models with Eric Easley and Adam Jermyn.

  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.

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 Tom Conerly yet. Send us a free-licensed photo.

  1. Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback arxiv.org. Checked 21 October 2026.
  2. Training a helpful and harmless assistant with RLHF anthropic.com. Checked 21 October 2026.
  3. Scaling Laws and Interpretability of Learning from Repeated Data arxiv.org. Checked 21 October 2026.
  4. Language Models (Mostly) Know What They Know arxiv.org. Checked 21 October 2026.
  5. Constitutional AI: Harmlessness from AI Feedback arxiv.org. Checked 21 October 2026.
  6. Towards Monosemanticity: Decomposing Language Models With Dictionary Learning transformer-circuits.pub. Checked 21 October 2026.
  7. Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet transformer-circuits.pub. Checked 21 October 2026.
  8. Sparse Crosscoders for Cross-Layer Features and Model Diffing transformer-circuits.pub. Checked 21 October 2026.
  9. Circuit Tracing: Revealing Computational Graphs in Language Models transformer-circuits.pub. Checked 21 October 2026.
  10. On the Biology of a Large Language Model transformer-circuits.pub. Checked 21 October 2026.
  11. A Toy Model of Interference Weights transformer-circuits.pub. Checked 21 October 2026.
  12. Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations transformer-circuits.pub. Checked 21 October 2026.
  13. Verbalizable Representations Form a Global Workspace in Language Models transformer-circuits.pub. Checked 21 October 2026.