PeopleofSuperintelligence

Researcher

Josh Abramson

Also known as Joshua Abramson

Josh Abramson is a machine learning researcher whose work spans interactive agents and AI for protein science. He earned a statistics PhD at UC Berkeley in 2012, then studied memory, imitation and human feedback as ways to train agents in virtual environments. His work includes the Multimodal Interactive Agent and equal first authorship of the 2024 AlphaFold 3 paper on biomolecular structure prediction. As of October 2026, he is a machine learning engineer and researcher at Google DeepMind, where he coauthors SynthIDBio, a method family for watermarking AI-generated protein sequences and structures.

Works
21
Citations
16,093
h-index
10

Publication numbers from OpenAlex, as of 15 October 2026.

Timeline

  1. 2012

    Statistics PhD at Berkeley

    He completes a Statistics PhD at UC Berkeley in 2012 with a dissertation on minorants and majorants of random walks and Lévy processes.

  2. 2018

    MERLIN predictive memory

    In 2018, he coauthors Unsupervised Predictive Memory in a Goal-Directed Agent. The MERLIN model uses predictive modeling to guide memory formation.

  3. 2020

    Imitating interactive intelligence

    In 2020, he coauthors Imitating Interactive Intelligence, a study of agents designed to interact naturally with humans in virtual environments.

  4. December 2021

    The Multimodal Interactive Agent

    In December 2021, he coauthors the Multimodal Interactive Agent (MIA), which combines vision, language, navigation and manipulation for extended interactions with people.

  5. November 2022

    Human feedback for agents

    In November 2022, he coauthors a study that uses reinforcement learning from human feedback to improve embodied agents in simulated 3D environments.

  6. 2024

    Equal first authorship of AlphaFold 3

    In 2024, he shares equal first authorship of Accurate structure prediction of biomolecular interactions with AlphaFold 3. The paper presents a diffusion-based model for predicting structures across biomolecular complexes.

  7. As of October 2026

    Structural biology at DeepMind

    In September 2026, he coauthors SynthIDBio, a family of methods that watermark AI-generated protein sequences and structures. As of October 2026, he is a machine learning engineer and researcher at Google DeepMind, focusing on structural biology.

Sources and credits

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

  1. Joshua Abramson | Department of Statistics live-statistics.pantheon.berkeley.edu. Checked 15 October 2026.
  2. Unsupervised Predictive Memory in a Goal-Directed Agent arxiv.org. Checked 15 October 2026.
  3. Imitating Interactive Intelligence arxiv.org. Checked 15 October 2026.
  4. Creating Interactive Agents with Imitation Learning deepmind.google. Checked 15 October 2026.
  5. Improving Multimodal Interactive Agents with Reinforcement Learning from Human Feedback arxiv.org. Checked 15 October 2026.
  6. Accurate structure prediction of biomolecular interactions with AlphaFold 3 nature.com. Checked 15 October 2026.
  7. Function-preserving watermarking of AI-generated proteins nature.com. Checked 15 October 2026.
  8. Trainer biographies | AlphaFold Education Summit ebi.ac.uk. Checked 15 October 2026.