Postdoc / CLE Moore Instructor, MIT

I'm interested in building foundational algorithms for studying AI. My work so far has produced new methods for tracing AI-generated outputs on the internet and for attributing model behavior to training data. These methods can be proven correct under natural and realistic assumptions---an outcome made possible not by understanding AI in all its complexity, but by carefully identifying simple abstractions that we can work with.
I'm open to new projects in AI safety that have the potential to yield satisfying and rigorous solutions.
I'm also open to continuing work on watermarking or data attribution, including:
- Building practical, undetectable watermarking schemes with semantic robustness.
- Making data attribution with provable correctness scale to large models and datasets.
- Studying theoretical limitations of potential data attribution methods.
Experience with research in theoretical computer science is beneficial.