I work on mechanistic interpretability of neural networks, primarily focusing on sparse autoencoders and extracting interpretable feature representations. I want to empirically understand and reason about models. I care about advancing the safety and steerability of AI systems but also find discovering their underlying structure intrinsically intellectually meaningful. I believe AI has overwhelming potential to make the future amazing and I want to do my part in ensuring we realise that future.

I am currently a member of technical staff (founding research scientist) at Goodfire. Previously, I’ve focused on a few different pursuits that can almost all be condensed down to “use computers to do good within biology”.

Contact Me!

I’m always excited about speaking to new people! My DMs are open on Twitter or you can email me liv[at]livgorton[dot]com.

I'm presently thinking about:

• Feature manifolds. Multidimensional features
• Symmetry in neural networks. How to think about symmetry in language models.
• Mechanistic interpretability for validation of chain-of-thought faithfulness.