👋 Hi! I'm Aarohi Zade.

I love AI/ML, data, cognitive science, art.

I'm about to graduate with a Master's degree in Analytics from UC Berkeley, and I hold a bachelor's degree from UC San Diego in Cognitive Science with a specialization in Machine Learning and Neural Computation.

Currently based in 📍 Berkeley, CA. Always happy to chat about science, art, development, and design.

A browser-based solitaire game inspired by a childhood favorite. Built with HTML deployed with Vercel.

*Research topic taken from Neel Nanda’s* [200 Concrete Open Problems in Mechanistic Interpretability](https://docs.google.com/spreadsheets/d/1oOdrQ80jDK-aGn-EVdDt3dg65GhmzrvBWzJ6MUZB8n4/edit?gid=0#gid=0)*. Huge thanks for all of the MechInterp resources!*In this project, I train a GPT-2 style transformer using rotary positional encodings with a single attention head. This basic transformer structure allows me to run attention experiments more simply, and to focus on the role of rotary embeddings instead of standard sinusoidal embeddings. I use the TransformerLens library to “peek” into the transformer, using visualizations of attention head activations.

Describe what this project is and what makes it interesting.

Describe what this project is and what makes it interesting.