Micah Carroll
Member Of Technical Staff Research Scientist at OpenAI
San Francisco Bay Area United States
Summary
👤
Senior🎓
Top SchoolMicah Carroll is a research-focused software engineer with 11 years of experience at the intersection of AI, human-AI coordination, and safety, now a Member of Technical Staff / Research Scientist at OpenAI. He helped build Overcooked-AI, a widely used benchmark for cooperative human-AI performance, and has published work on how deep RL agents fail to coordinate with humans. His background spans academic research stints at Cambridge and CHAI, an internship at Microsoft, and teaching roles at UC Berkeley, blending rigorous empirical work with engineering skills in test automation and reproducible benchmarks. Micah’s long-term focus is ensuring widespread AI benefits align with human values, and he brings a rare combination of hands-on QA/engineering experience and principled safety research.
11 years of coding experience
4 years of employment as a software developer
PhD Artificial Intelligence, PhD Artificial Intelligence at University of California, BerkeleyGraduated by completing the Esame Di Stato Classical and Ancient Studies, Graduated by completing the Esame Di Stato Classical and Ancient Studies at Liceo Classico ISIS Niccolini Palli, Livorno, Italy
English, Italian, Latin, Greek, German
Github Skills (9)
unit-testing10
python10
unit-test10
test-automation10
reinforcement-learning9
machine-learning9
artificial-intelligence9
deep-learning9
pytorch9
Programming languages (6)
Github contributions (5)
HumanCompatibleAI/overcooked_ai
Jun 2019 - Jan 2023
A benchmark environment for fully cooperative human-AI performance.
Role in this project:
QA Engineer / Test Automation Engineer
QA Engineer / Test Automation EngineerContributions:68 reviews, 242 commits, 62 PRs in 3 years 7 months
Contributions summary:Micah implemented test files and test suites for the overcooked_ai environment, specifically focusing on agent behaviors and the OvercookedGridworld MDP. The commits demonstrate the creation of test cases for various agent types, including FixedPlanAgent, CoupledPlanningAgent, and GreedyHumanModel, and the testing of scenario-specific behaviors within the environment. The code changes include the addition of test fixtures, assertions, and the use of the unittest framework.
pytorchhuman-aideep-learningreinforcement-learningmachine-learning
marcus-jw/Targeted-Manipulation-and-Deception-in-LLMs
Jun 2024 - Oct 2024
Codebase for "Targeted Manipulation and Deception Emerge in LLMs Trained on User Feedback". This repo implements a generative multi-turn RL environment with support for agent, user, user feedback, transition and veto models. It also implements KTO and expert iteration for training on user preferences.
Contributions:245 reviews, 107 PRs, 651 pushes in 3 months
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.