InfoQ 中文T2·78 pts
- Discovery Loop has built its own agent harness and large-scale multi-agent infrastructure, running about 20 automated improvement loops
- The team draws researchers from Gemini, DeepSeek and Kimi projects, with roughly 400 interviews conducted in its first three weeks
- Fei-Fei Li is skeptical of unbounded AI self-acceleration, stressing that humans must decide collective goals