Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
ASH: Agents that Self-Hone via Embodied Learning
TL;DR
Long-horizon embodied tasks in AI struggle due to reliance on hand-engineered rewards and expert demonstrations. ASH is introduced as an agentic system that learns an embodied policy from unlabeled internet video, using a self-improvement loop.
✦ Why It Matters
ASH enables engineers to create AI systems that learn autonomously from unstructured data, reducing reliance on manual input.
Key Takeaways
- Explore unsupervised learning techniques for AI training.
- Utilize long-term memory retention in AI systems.
- Implement self-improvement mechanisms in agent design.
- Evaluate AI performance in complex, multi-hour tasks.
How It Works
ASH operates by analyzing its own actions and outcomes to build an Inverse Dynamics Model (IDM), which helps it understand the relationships between actions and states. This model allows ASH to extract relevant information from large datasets of unlabeled videos, identifying key moments that can inform its decision-making process.
By retaining these moments as long-term memory, ASH can improve its performance over time, particularly in tasks that require extensive planning and strategy.