Rhythm Garg is co-founder and chief technology officer of Applied Compute, where he develops reinforcement-learning systems that help companies train specialized models on their own work. A former OpenAI researcher and core contributor to OpenAI o1, he is applying techniques behind frontier reasoning models to enterprise-specific problems, evaluations, and production feedback.
Garg studied at Stanford before joining OpenAI, where he worked on reinforcement learning and machine reasoning. He contributed to reasoning research for the o1 model and co-authored research on competitive programming with large reasoning models, contributing specifically to test-time strategy.
He founded Applied Compute with fellow former OpenAI researchers Yash Patil and Linden Li, becoming CTO alongside Patil as CEO and Li as chief architect. Garg calls their approach Specific Intelligence: custom models and agents trained against a company’s proprietary tasks and continually improved through real-world usage. In his announcement of the company, he named Cognition, DoorDash, and Mercor among its customers or collaborators.
- Private benchmarks and production feedback. Garg measures enterprise AI against organization-specific tasks, using deployment data and user feedback to keep specialized models improving after they enter production.
- Asynchronous reinforcement learning. In work presented with Li, Garg explained how synchronous training leaves expensive hardware idle while unusually long responses finish. Overlapping sampling and training increases utilization, but creates policy staleness when responses incorporate older model weights; excessive staleness increases variance and can destabilize learning.
- Predictable enterprise training. Garg treats consistent completion times, affordable compute, and stable training as product requirements. His goal is reinforcement-learning infrastructure capable of delivering specialized customer models quickly without sacrificing reliable deployment or sustainable unit economics.
