Girish Gupta

I'm an AI interpretability researcher working to understand how artificial intelligence models actually work. I believe mechanistic interpretability is essential for AI safety—making it one of the most important scientific challenges of our time.

My approach draws on an unusual combination: theoretical physics (my academic foundation), investigative journalism (a decade reporting on global crises), and Silicon Valley engineering (building AI systems across multiple industries).

As an engineer and tech leader, I built AI infrastructure and software across news, public policy, human rights investigations, accounting, and healthcare, as well as software to bring simple, live economic data to people suffering hyperinflation. I've seen firsthand how shortcuts and misunderstandings can lead to serious real-world consequences.

Before entering tech, I reported on global crises—primarily Venezuela, where I lived for nearly a decade, as well as Iraq, Afghanistan, Colombia, Cuba, Mexico, Lebanon, Jordan, Guyana, Brazil, and Egypt—for the New Yorker, New York Times, Reuters, Time Magazine, the BBC, NPR, and many others. This led me to write a memoir, Always Go, which is a sharp critique of a news industry shaped often similarly to AI by reward hacking and misaligned incentives.

In my spare time, I enjoy photography, learning history, and playing the piano. I live in San Francisco with my wife and children.

Currently: Taking BlueDot's Technical AI Safety course, mentoring interpretability research through Algoverse, and building open-source tools for mechanistic interpretability. I'll be at NeurIPS in San Diego later this week. Please email or message me if you'd like to meet!