Jonas Schneider, Head of Engineering for Robotics, OpenAI
summary
OpenAI builds end-to-end ML infrastructure that runs on Kubernetes across Azure/AWS and private datacenters, mixing TensorFlow with custom ops, MPI for sync and Redis for parameter exchange, and CI-style regression tests. Evolution Strategies cut a robot walking training job from about one day to ~10 minutes. A Dota agent self‑plays to surpass top humans in weeks. Focus is on orchestration and observability, measured metrics (TrueSkill) and practical engineering tradeoffs.
AI-written summary · about
chapters (24)
- 0:00Introduction
- 0:37What is OpenAI
- 1:09Robot in VR
- 1:26Playing Dota 2
- 1:56Machine Learning
- 2:26Machine Learning Core
- 3:11Kubernetes
- 3:32Cloud Switching
- 4:05GPU Instances
- 4:39Declarative Infrastructure
- 5:36Research vs Engineering
- 6:04Algorithms
- 6:51Other Axis
- 7:20RL Algorithms
- 8:16Reinforcement Learning
- 9:27Training Topology
- 10:08Scaling Your Team
- 11:27The Usual Scientific Approach
- 12:01Limitations
- 12:49Mix and match
- 14:03Metrics
- 15:39Regression testing
- 17:14Trade off
- 18:03Summary
related talks
28mScaling AI Infrastructure
14mKeynote: Building the Infrastructure that Powers the Future of AI
35mHow The AI Institute is Revolutionizing Robotics ML Training
13mAI Agents for Autonomous Industrial Operations on
32mThe Road to General AI
18mAI Owes its Rebirth to Open Source and the Cloud
18mRobotics: why now?
26mEngineering Excellence: Platforms and Architectures for Real-Time AI Analytics