Tommy He
I'm interested in understanding the principles of intelligence to design more robust systems including:
- Interpretability: mech interp, representation science, data attribution
- Open-ended learning: self-play, open-endedness, alife
Previously, I was a YC founder building Clarum, AI agents for private capital investing. I've also worked on high frequency trading at Tower Research and privacy preserving knowledge intersection at Optable.
Research
-
Context–Answer Metamodels: Forecasting LLM Activations Before the
First Answer Token
(2026)
Tommy He, Thomas Jiralerspong
NeurIPS 2026 NeuralArtifacts (oral spotlight)
[paper] [code] -
tufazip (2026)
Tommy He
Large Text Compression Benchmark (#2)
[blog] [code] [benchmark] -
A Predictive Law for On-Policy Self-Distillation From World
Feedback
(2026)
Tommy He, Jerome Sieber*, Matteo Saponati*
ICML 2026 RLxF
[arxiv] [code]