I’m a Computer Science PhD student at UC Berkeley working on how to align powerful future AI systems with human values, to reduce the existential risk posed by AI. I am advised by Stuart Russell and part of the Center for Human-Compatible AI. I am grateful to be supported by fellowships from the Future of Life Institute and Open Philanthropy.
PhD in Computer Science, since 2022
MSc in Artificial Intelligence, 2022
University of Amsterdam
BSc in Physics, 2020
einsumis one of the most useful functions in Numpy/Pytorch/Tensorflow and yet many people don’t use it. It seems to have a reputation as being difficult to understand and use, which is completely backwards in my view: the reason
einsumis great is precisely because it is easier to use and reason about than the alternatives. So this post tries to set the record straight and show how simple
We prove impossibility results showing that Karger’s contraction algorithm cannot be extended to $s$-$t$-mincuts or normalized cuts. However, we show how extensions of Karger’s algorithm can still be useful for seeded segmentation.
We present a method for simplifying a learned reward model before visualizing it and show that this can make the reward more interpretable.