Jacob Steinhardt

Also known as Jacob S. Steinhardt

Jacob Steinhardt is a machine-learning researcher who works on making AI systems understood by and aligned with humans. A Hertz Fellow from MIT, he coauthored Concrete Problems in AI Safety in 2016 and completed a Stanford PhD on robust learning under Percy Liang in 2018. After a post-doctoral year at OpenAI and Open Philanthropy, he joined Berkeley's Statistics faculty in 2019, and in 2020 he helped create MMLU, which became the most common benchmark for measuring AI capabilities. In October 2024, he and Sarah Schwettmann launched Transluce, a non-profit lab that builds open tools for understanding AI systems. Today he is an associate professor of statistics and EECS at UC Berkeley and the CEO of Transluce.

Timeline

By 2012

MIT and a Hertz Fellowship

As an undergraduate mathematics student at MIT, he starts to think about a future in which AI automates most cognitive work. He becomes a Hertz Fellow in 2012 and goes on to graduate study in artificial intelligence at Stanford.
June 2016

Concrete Problems in AI Safety

In June 2016, he coauthors Concrete Problems in AI Safety with Dario Amodei, Chris Olah, Paul Christiano, John Schulman and Dan Mané. The paper sets out five practical research problems about accidents in machine learning systems, among them reward hacking, safe exploration and distributional shift.
2017–2018

Robust learning and a Stanford PhD

In 2017, he publishes Learning from Untrusted Data with Moses Charikar and Gregory Valiant at STOC, and Certified Defenses for Data Poisoning Attacks with Pang Wei Koh and Percy Liang at NIPS. He completes his Stanford PhD under Liang in 2018 with the dissertation Robust Learning: Information Theory and Algorithms.
2019

OpenAI, then Berkeley Statistics

He spends his post-doctoral year working at OpenAI and the Open Philanthropy Project. In fall 2019, he joins the Statistics faculty at UC Berkeley as an assistant professor, with membership in the EECS department and the Berkeley Artificial Intelligence Lab. He goes on to teach Stat 165, Berkeley's course on forecasting.
September 2020

The MMLU benchmark

In September 2020, he and Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika and Dawn Song introduce Measuring Massive Multitask Language Understanding. MMLU, a test of 57 tasks ranging from elementary mathematics and US history to law, becomes the most common benchmark for measuring the capabilities of AI systems.
2023

An AI2050 Early Career Fellowship

He becomes a 2023 AI2050 Early Career Fellow. His project aims to build AI systems that discover and explain important facts about the behavior and possible failure modes of other AI systems.
October 2024

Co-founding Transluce

In October 2024, he and Sarah Schwettmann, a research scientist at MIT CSAIL, launch Transluce, a non-profit lab building open, scalable technology to understand AI systems. Its early tools include Monitor, an interface for understanding and guiding the internal computations of large language models. He presents on AI risks and rewards at the Hinton Lectures in October 2024.
As of October 2026

Associate professor and Transluce CEO

He is an associate professor of statistics and EECS at UC Berkeley, where he is part of BAIR and CLIMB, and founder and CEO of Transluce. His research focuses on making machine learning systems understood by and aligned with humans. For the Summer 2026 MATS program, he and Schwettmann mentor scholars on interpretability tools, behavior elicitation and LLM judges for Transluce's Docent.