Building a Go AI with Kubernetes and TensorFlow
summary
An open‑source AlphaZero rewrite called MiniGo trains by massive self‑play and runs the pipeline on Kubernetes with TensorFlow, storing models and SGFs in GCS. Engineering choices: containerized workers launched as 20,000 parallel self‑play jobs that pull the latest model, write games, and exit; switching to NVIDIA GPUs and CUDA drivers cuts inference time from seconds to hundreds per second; Docker images plus GKE scale the workload and the demo shows working models despite known ladder weaknesses.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:51What is MiniGo
- 2:27Building MiniGo
- 3:19MiniGo Demo
- 4:07What is Go
- 12:29Reinforcement Learning Loop
- 14:32Recap
- 21:47Demo
- 26:19Wrapup
- 31:14TensorFlow and Go
related talks
37mA Funny Thing Happened On The Way To Reimplementing AlphaGo
38mAI in Kubernetes: How to Get Started?
29mThe Go-to Language for AI: Opportunities and Challenges
27mHarnessing Kubernetes for Scalable AI/ML Workloads
04mTensorFlow and Go
20mUsing TensorFlow with Go
31mDecoding AI: A Go Programmer's Perspective
50mDeep Learning with Go