ChatGPT and Minecraft's diamonds playing agents: Reinforcement Learning in a nutshell with sheeprl
summary
Sheeprl is an open source Deep RL framework that runs PPO agents, uses distributed PyTorch and Lightning Fabric for multi-CPU/GPU training, and ships examples that train Pong and a Dreamer agent that crafts diamonds in Minecraft. The talk cites AlphaTensor finding a 4x4 matrix multiply that drops 49→47 multiplications, explains Reinforcement Learning Human Feedback as a three‑stage RLHF pipeline, reports Dreamer with a 15‑step horizon and ~2 days on 2 GPUs, and shows two architectures: coupled parallel workers and a decoupled env/train setup.
AI-written summary · about
chapters (4)
related talks
01h 09mMinecraft as AI Playground and Laboratory
30mReinforcement Learning - ChatGPT, Playing Games & More
41mReinforcement Learning and Minecraft
31mDesigning game AI with reinforcement learning
28mAI and Minecraft Lars
50mIEEE CoG 2019 - Industry Day - Katja Hofmann (Microsoft Research)
32mAI and Minecraft
21mAdvanced Ray RLlib: New Features and New Possibilities