Google's Gemini model demonstrated a form of recursive self-improvement by co-designing an evolutionary algorithm that sped up its own training by 1%.

The Center for Security and Emerging Technology (CSET) held a workshop with participants from frontier model developers who failed to establish a consensus expectation about the impact of automated AI...

Independent audit requirements have been included in several AI policy proposals but are repeatedly stripped out due to industry lobbying.

Helen Toner argues that if automated AI R&D significantly speeds up progress at closed labs, it would challenge the assumption that open-source models will always remain close behind in capability.

Participants in the CSET workshop on automated AI R&D found they largely agreed on developments expected by 2026-2027, but this agreement makes it difficult to identify in advance whether a recursive ...

One skeptical view of rapid AI-driven R&D acceleration is based on Amdahl's Law, which suggests that speeding up one part of the process, such as coding, will cause bottlenecks to appear elsewhere, li...