Will this project in mechanistic interpretability make me happy by the end of 2024?

By the end of 2024 will I think that pursuing the following research project was a good idea?

The project: take an equivariant graph neural network or a similar architecture designed for learning to solve physics problems; train it to predict some non-trivial simulated dynamics (e.g. the gravitational three body problem); do mechinterp to find out what exactly the deep learning model has learned.

I will be happy about this if the project succeeds to a degree that I feel comfortable submitting a paper to a physics or astrophysics journal or an extended abstract to some ML conference (including workshops). In that case I will resolve YES. If I will not pursue the project at all (e.g. because of leaving academia) I will resolve NA. If I instead fail to obtain sufficiently interesting results despite spending some effort on this I will resolve NO.

This is subjective so I will not bet. I expect some details of the project to change along the way (if we knew what we were doing it wouldn’t be called research).

Get Ṁ600 play money

More related questions