Visualizing Hierarchical Reasoning Model training on a BabyAI task

What does an AI actually learn during training? Most explanations skip this part.

This video opens the black box of AI training. Watch a hierarchical reinforcement learning model transform from random wandering to efficient problem-solving on the classic BabyAI task.

We explore:
The BabyAI unlock-and-open task (navigate to key, pick it up, unlock door)
Hierarchical Reinforcement Machines (HRM) with planner and doer agents
Real-time visualization of the learning process
How thought bubbles reveal what the AI is "thinking"

Try the interactive visualization yourself and watch AI learning happen in real-time.

LINKS
Interactive Visualization: https://github.com/softwarewrighter/v...
GitHub Repo: https://github.com/softwarewrighter/v...
HRM Paper: https://arxiv.org/abs/2506.21734
TRM Paper: https://arxiv.org/abs/2510.04871
BabyAI Paper: https://arxiv.org/abs/1810.08272

TIMESTAMPS
0:00 Intro
0:05 The problem with AI training explanations
0:20 What this visualization shows
0:39 The BabyAI task explained
1:10 Why hierarchical learning?
1:48 Planner and doer roles
2:27 HRM vs LLM comparison
2:54 The visualization walkthrough
3:30 Training data and learning process
4:06 Interactive demo
6:19 Key takeaways
6:33 Try it yourself

#MachineLearning #ReinforcementLearning #AIVisualization #BabyAI #HierarchicalRL #AITraining #DeepLearning #vibecoding #AIExplained #LearnAI