Co-Founder and Chief Architect at Applied Compute, building production-grade reinforcement learning infrastructure for training AI agents at scale. Previously worked at OpenAI on reasoning models, coding AI, and deep research tools including the o1 model. Stanford graduate (2021-2023) specializing in ML systems and infrastructure.
From OpenAI to Applied Compute
After working on infrastructure and ML systems for reinforcement learning training at OpenAI, Linden Li co-founded Applied Compute in 2025 alongside Rhythm Garg and Yash Patil. The company has raised $80M+ in funding (backed by Benchmark, Sequoia, Lux, Hanabi, Neo) and is valued at close to $700M.
Applied Compute’s mission: building “Specific Intelligence” to create the first generation of agent workforces through efficient, production-ready RL infrastructure.
Technical Philosophy
Linden’s approach challenges academic RL by prioritizing practical constraints:
Pipeline RL Architecture - Async RL with in-flight weight updates enables faster training. The key innovation: managing staleness (where tokens are sampled from previous weight versions) through training stability advances.
First-Principles Modeling - High-throughput RL requires balancing GPU count, training batch size, KV cache management, sampling throughput, and training throughput. Too many training operations without enough sampling (or vice versa) creates inefficiency.
Production Constraints - Fast, cheap, and predictable training enables organizations to build domain-specific agents that outperform frontier models on specialized problems with reasonable resources.
Summit Session
Conference: AI Engineering Code Summit 2025 Date: November 21, 2025 Time: 11:20 AM - 11:39 AM Topic: Efficient Reinforcement Learning (co-presented with Rhythm Garg)
The session explored how to push AI past productivity into measurable value through RL mechanisms. Key insights included:
- High-compute RL helps LLMs learn to reason by running models hundreds of times, grading answers, and reinforcing correct thinking paths
- Pipeline RL makes this approach viable at production scale through async training and staleness management
- The “data flywheel” enables continuous improvement of agents through deployment and learning
Background
- Education: Stanford University (2021-2023), Computer Science
- Previous Role: OpenAI - Infrastructure and ML systems for RL training, worked on reasoning models (o1), coding AI, and ChatGPT deep research tools
- Recognition: World Schools Debating Champion (2018), Frederick Emmons Terman Engineering Scholastic Award (2023)
- Location: San Francisco Bay Area
Industry Impact
Applied Compute’s infrastructure represents a paradigm shift from “research labs only” to production-grade RL. Organizations no longer face a binary choice between using frontier models as-is or training from scratch—they can create specialized agents with reasonable computational resources.
Last Updated: 2025-11-24