Aayan Rizvi
I'm Aayan, a third-year at UC Berkeley studying computer science and data science, with a concentration in applied math. I'm broadly interested in everything ML. Right now that mostly means mechanistic interpretability and video generation.
I work on empirical game theory in the Strang Lab. Before that, I did cost optimization at Bespoke Labs, research on video and multimodal understanding with Autodesk, and LLM post-training and document-analysis agents at Roche. I'm currently a TA for Data 140.
Things I'm thinking about
- Reading networks one neuron at a time: when an explanation of a neuron is actually enough to replace it
- Weight-space programming: old ideas about writing weights directly, revisited with LoRAs and modern generative models
- What optimizers teach: whether AdamW and Muon learn different representations, beyond their loss curves
Selected projects
All projects →- Neural networks are kinda clean2026
Can you read a language model one neuron at a time? I replaced Qwen3-4B neurons with simulations built from their own explanations and measured what broke.
- Adversarial autoresearch for GLEE2026
My agent for GLEE, a NeurIPS 2026 competition on LLM economic games. An LLM research operator proposed every change, and throwaway critic agents had to fail to kill it before it shipped.
Writing
All writing →Coursework
- CS 189Machine Learning
- CS 182Deep Neural Networks
- CS 180Computer Vision
- CS 188Artificial Intelligence
- CS 170Algorithms
- Data 140Probability
- Math 156Numerical Analysis
Recently watched
Letterboxd →- Incantation 2022★★review
- You Can See Everything 2026★★★★★review
- Primal Fear 1996★★★½
- 22 Jump Street 2014★★★½