San Francisco · Intology
I build AI systems that do science.
I'm a co-founder of Intology, a research lab automating the process of discovery. We build Artificial Scientists: systems that read the literature, form a hypothesis, run the experiments, and write up what they found — with no human in the loop.
Two exist so far. Zochi was the first AI system to pass peer review at a top-tier venue. Locus is the first to outperform human experts at AI R&D.
Before Intology I did research at Meta and Virtue AI, and founded Lapis Labs, a student-led group that published 20+ papers with collaborators including Google DeepMind, the Center for AI Safety, and IBM Research.
Artificial Scientist for AI R&D
2025 —
A long-horizon research system that runs its own experiments for days at a time. Locus is the first AI system to beat human experts on RE-Bench given the same time and compute, and in January 2026 it set a world record on the human NanoGPT Speedrun leaderboard with a fused Triton kernel it designed, implemented, and debugged itself.
The first Artificial Scientist
2025
Zochi takes a research question from literature review through experimentation to a written paper. Its paper Tempest was accepted to the main proceedings of ACL 2025 — the first fully AI-generated discovery to clear peer review at an A*-ranked venue, scoring in the top 8.2% of submissions by meta-review score.
Measuring real research ability
2026
Benchmarks for research agents are easy to saturate and easy to contaminate. NanoGPT-Bench drops agents into the GPT-2 pretraining speedrun at a fixed human world record and asks them to make it faster, with no internet and no human help. Frontier coding agents recover less than 10% of what the human community achieved over the following five months, spending most of their compute on hyperparameter tuning instead of the algorithmic work that actually moves the record.
Agents, robustness, and AI safety, from before Intology. The complete list lives on Google Scholar.
The first search algorithm for language agents, combining reasoning, acting, and planning into one tree search. Now implemented in LangChain and LlamaIndex.
A defense objective and an algorithm for finding trigger tokens that enforce harmless behavior, transferring across jailbreaks and models.
5,694 prompts spanning 314 risk categories drawn from 16 company policies and 8 government regulations.
Domain knowledge as logical constraints over an ensemble of specialized models, deployed for anomaly detection on production graph data.
Teacher gradients generate hard augmentations during distillation, yielding the most OOD-robust ResNet34 and ResNet50 of their time.
May 2026
We released NanoGPT-Bench, an evaluation of whether coding agents can actually do research. The best one recovers 9.3% of five months of human progress.
Jan 2026
Locus set a world record on the human NanoGPT Speedrun leaderboard with a fused Triton kernel it designed on its own. Later human records built on top of it.
Nov 2025
We introduced Locus, the first AI system to outperform human experts at AI R&D.
May 2025
Zochi published at ACL 2025 Main — the first AI system to independently pass peer review at an A* venue.
Mar 2025
We launched Intology and published the Zochi technical report.
Sep 2024
Three papers at NeurIPS 2024, with RPO accepted as a Spotlight (top 3%), and our AIR-Bench work was covered by Wired.
May 2024
LATS was accepted at ICML 2024 and is now implemented in LangChain and LlamaIndex.