JAY ZENITH

[github] [x]

I built PREDICT to understand the post-training stack end to end, while testing whether an agent could learn to predict the consequences of its own actions before observing the result, using coding as the testbed.

Now I’m most intrigued by self-improving agent systems: how they could take on more of the research required to improve their successors, develop better research judgment, coordinate and explore across many trajectories, and remain understandable and aligned as that capability compounds.

  • predict

    Reactive vs. Predictive Post-Training for Coding Agents. I trained two Qwen3-4B coding agents on the same verifier-backed environment using Prime Intellect's Verifiers and PRIME-RL: reactive test-and-recover with GRPO versus predictive KEEP/REVISE with GRPO + auxiliary CE. Both improved with RL, but the predictive arm did not outperform. Trajectory analysis showed why: rollouts that chose REVISE almost never recovered, while cheap test feedback made direct observation more useful than foresight. [read the study] [code]