Today we are introducing BlockAssist, an AI assistant that learns from its user’s actions in Minecraft. The assistant appears in-game with you, starting with only basic knowledge of the game’s commands. As you play, it learns how to assist you in building, learning directly from your actions. We’re open sourcing it today as an early demo of assistance learning - a new paradigm for aligning agents to human preferences across domains.

‍

Background

Assistance learning lets agents learn directly from human actions instead of hand-crafted reward functions. Because preference data is captured automatically, it scales better than reinforcement learning from human feedback (RLHF), which relies on manual labeling. Early research in the field has focused on games, but the technique applies to any domain that requires human preference tuning or lacks clear, verifiable rewards. We’re demonstrating the concept in Minecraft today, but will soon show how it applies more broadly as well.

‍

How it Works

BlockAssist follows the setup introduced in the AssistanceZero paper. The human player and the AI assistant share a hidden goal (e.g. building a house), but only the human knows what that goal is. The assistant must infer the goal on-the-fly by predicting human actions and rewards with a neural network and Monte-Carlo Tree Search (MCTS). After the user completes an episode, the recording is used to train an updated assistant model, which is then uploaded to Hugging Face. User progress is tracked on the global Gensyn leaderboard.

‍
Why it Matters

BlockAssist demonstrates a new way to train agents - directly from human actions rather than manual labels. This creates higher signal datapoints and runs passively while users are engaging in the relevant activity. And because training runs on Gensyn’s permissionless compute network, anyone can contribute gameplay data and compute, accelerating progress without a central gatekeeper.

‍

Get Involved

1. Play BlockAssist today and begin training the best assistant you can. Every block you place makes the assistant smarter. 
2. Share your progress with the community by posting your gameplay videos, stats, and Hugging Face uploads. Track your participation on the leaderboard.
3. Stay tuned for more updates as we expand assistance learning into new domains.  

‍