Linden Li is co-founder and chief architect of Applied Compute, which builds specialized enterprise agents that learn from an organization’s own data, workflows, and performance requirements. Previously a researcher at OpenAI, he combines reinforcement learning, inference engineering, and model-training infrastructure to make company-specific intelligence economically practical.
From computer vision to frontier-model infrastructure
At Stanford, Li worked with Fei-Fei Li on interpretable video understanding before research internships at NVIDIA and MosaicML deepened his interest in efficient model training and distributed systems.
In 2022, he and Abhi Venigalla estimated that optimized infrastructure and compute-aware training could produce GPT-3-level model quality for less than $500,000. Their figure was a projection based on scaling laws and infrastructure benchmarks, not the cost of a completed training run. At MosaicML and Databricks, Li also co-authored practical guidance on inference latency, batching, memory bandwidth, and throughput.
His interests extended to controllable image generation. He co-authored Collage Diffusion, a layered approach that lets users arrange objects while a diffusion model reconciles their lighting, perspective, and composition, and helped build its open-source interface.
At OpenAI, Li worked on machine-learning systems and reinforcement-learning infrastructure, contributing to the o1 system card and scaling work behind GPT-4.5. He subsequently founded Applied Compute with fellow OpenAI alumni Rhythm Garg and Yash Patil. The company has reported $160 million in total funding and customers including Cognition, DoorDash, and Mercor.
- Specific Intelligence: Applied Compute trains agents around company-specific tasks, institutional knowledge, and measurable business outcomes, then improves them through operational feedback.
- Asynchronous reinforcement learning: In his AI Engineer Code session with Garg, Li modeled how to divide a fixed GPU budget between sampling and training while accounting for response-length variability, memory limits, and stale rollouts. One simulated configuration projected a roughly 60% speedup over synchronous training; the figure describes that modeled scenario, not guaranteed production performance.
- Leverage thresholding: In research with Agastya Goel, Li argues that rare successful attempts can carry disproportionately valuable learning signals. Prioritizing informative samples frees compute for generating additional attempts.
- Agentic inference benchmarking: With Oam Patel, Li analyzed production-like agent workloads involving tool use, long contexts, and variable response lengths. Their open-source trie benchmarking harness tests how excessive concurrency and key-value-cache eviction can reduce useful throughput.
