I spend a lot of time thinking about learning and computation: how much can we improve by rethinking the algorithms and systems underneath? What flawed priors in the systems of today will be discarded in the systems of tomorrow? Other interests include: trying to make sense of the economy, investing in said economy, businesses, consciousness, correctness, and human motive.
I'm currently studying computer science & electrical eng at UC Berkeley. I'm also a research intern at Q Labs, focusing on efficient LLM pretraining. This past summer, I post-trained various agents at Standard Kernel Co. Previously, I was researcher #3 at AfterQuery, and interned at AMD on AI tooling. I'm part of Pear Prime 002.
At Berkeley Sky Lab, I've been fortunate to collaborate with Mert Cemri on ML systems. Before this, I worked with Lisa Dunlap at Berkeley AI Research on model behavior and safety.
In my spare time, I enjoy playing golf, basketball, chess, and (occasionally) poker. Please email me if you would like to chat :)
Latest
- Intuition as Compressed Verification
- Atlas for ML Research
- Hadamard and Monarch: Compressing GPT-2 Small
- Moravec's Paradox
- Autoencoders
- Is EAGLE-3 Speculative Decoding Lossless?
- From QK-Clip to Muon-Clip
- Parallel Transformer Blocks
- Token Embeddings Violate the Manifold Hypothesis
- Reflections on BayLearn
- Our project, Dementor, was accepted to BayLearn 2025 as an oral presentation (<5%)!
- Representative Sampling of High-Dimensional Point Sets
- We release VADER, a security benchmark for LLMs!
