Nova DasSarma

Timeline

1. By 2022

   Information systems at UMBC

   At 12, she works on bioinformatics while helping her microbiologist parents, and she also does comparative genomics of haloarchaea. By 2022, she has studied information systems at the University of Maryland, Baltimore County, worked as a system administrator for university research teams, and become a data engineer at various tech companies.

2. December 2021

   A language assistant paper

   The December 2021 paper A General Language Assistant as a Laboratory for Alignment lists her among its authors, alongside Amanda Askell, Dario Amodei and Jared Kaplan. It studies prompting and ranked preference modeling as baseline techniques toward a helpful, honest and harmless text assistant.

3. February 2022

   Co-founding Hofvarpnir Studios

   Hofvarpnir Studios, the nonprofit she cofounds with a grant from the Effective Altruism Infrastructure Fund, receives its tax exemption in February 2022. It aims to build a GPU cluster that academic researchers can use to improve the safety of cutting-edge AI systems, and she serves as its treasurer.

4. March 2022

   Core infrastructure on induction heads

   In March 2022, the Transformer Circuits Thread publishes In-context Learning and Induction Heads, an Anthropic paper that lists her among its Core Infrastructure Contributors. The paper frames mechanistic interpretability, the reverse engineering of what a model computes, as a way to address safety problems in Transformer language models.

5. April 2022–December 2022

   RLHF and Constitutional AI papers

   She co-authors Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback in April 2022, Language Models (Mostly) Know What They Know in July 2022, and Constitutional AI: Harmlessness from AI Feedback in December 2022. The first applies RLHF to finetune language models as assistants, the second tests whether models can predict which questions they can answer, and the third trains a harmless assistant without human labels for harmful outputs.

6. June 2022

   Lead systems architect at Anthropic

   In June 2022, she is Anthropic's lead systems architect, responsible for information and computer security and for engineering the company's computer systems. Her work includes securing the compute power for large language model experiments. In her view, infrastructure on the safety side is a limiting factor, and she aims to give safety researchers a differential boost.

7. January 2024

   The Sleeper Agents paper

   In January 2024, Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training appears on arXiv with her among its 39 authors. The paper builds proof-of-concept deceptive language models and finds that their backdoor behavior persists through supervised fine-tuning, reinforcement learning and adversarial training.

8. 2025

   Anthropic's compute on Amazon EKS

   At AWS re:Invent 2025, she joins a session on Amazon EKS as Member of Technical Staff and infrastructure lead at Anthropic. Her team operates more than 99% of Anthropic's compute on EKS, at a scale that has shaped both Amazon EKS Ultra-Scale Clusters and the Provisioned Control Plane feature.