- ByteDance's Seed AI team found that AI agents can double their learning speed every three months through interaction with real-world environments.
- The AI industry is shifting from brute-force data and compute scaling to agentic AI, but post-deployment learning in real-world settings is poorly understood.
- To study this, ByteDance created EdgeBench, a benchmark with 134 long-term tasks requiring over 12 hours of continuous AI agent operation across various domains.