Kevin Zhu

I founded Algoverse, where we mentor ambitious students into published ML researchers. I think about AI safety, mechanistic interpretability, and how to build institutions that produce great research.

Previously: quantitative research at Citadel, software engineering at Palantir. CS from UC Berkeley, where I was a 4x head TA for CS 170 and a researcher in the RISE Lab with Ion Stoica.

Always happy to chat — say hi!

now

Algoverse current

Founded an AI research mentorship program — students publishing at NeurIPS, EMNLP, and other top venues.

Mechanistic Interpretability

Understanding what language models are actually computing.

AI Safety

Cofounded the Algoverse AI Safety Fellowship, funded by Open Philanthropy.

selected work

Semantic Self-Consistency NeurIPS '24 MATH-AI
Featured in OpenAI's PaperBench — one of 22 papers selected to benchmark AI research replication.
Scratchpad Thinking (Spotlight) NeurIPS '25 Mech Interp
Latent reasoning models alternate between storage and computation across latent steps.
NovelHopQA EMNLP '25
Diagnosing multi-hop reasoning failures in long narrative contexts.

writing

A gentle introduction to algorithms — divide and conquer, FFT, graphs, DP, LP, and P/NP
A primer into UARS — the sleep disorder no one tests for
A primer on lucid dreaming — how I trained myself to dream on command