Jason Lu

The story

Four chapters

Teaching AI to read science, to live with people, to improve itself, and what comes next. Each scene comes to life as you scroll.

1. 1

   01 · DP Technology · Mar 2025 – Feb 2026

   Reading

   My first chapter was teaching machines to read science. On the Uni-Parser team at DP Technology I worked on the inference system that turns papers and patents, with their text, tables, formulas and molecules, into clean training data for models, and on MolParser-Mobile, a tiny model that reads molecule drawings.

   • Document AI
   • Inference at scale
   • Small, fast models
2. 2

   02 · Intent · Feb – Aug 2026

   Living with people

   Next I wanted AI to live with people, not just read about them. At Intent I built a team agent for group chats: one shared teammate that anyone in the group can @. It remembers the group, uses tools, runs long tasks in the background and follows up on its own, on a runtime built to recover from failures.

   • Multiplayer agents
   • Memory & tools
   • Long-running tasks
3. 3

   03 · Evolvent AI · May 2026 – now

   Improving themselves

   Then the bigger question: can agents improve AI itself? At Evolvent AI I co-lead RSIGym, an environment where frontier agents run AI research end to end. They synthesize data, train models, rewrite their own harnesses and are scored in sandboxed evals, round after round. I also co-designed RSI-Index, the score that tracks how far they get.

   • RL environments
   • Evals
   • Self-improving agents
4. 4

   04 · Columbia University · New York

   Now

   Now I am in New York for an M.S. in Computer Science at Columbia, after a B.S. at NYU Shanghai. Next I want to build this at a frontier lab: the environments, evals and training systems that let AI agents improve AI systems.

   • Research infrastructure
   • RL for agents
   • Frontier labs next