Anders Sandberg
This is an interesting overview, especially together with the paper https://arxiv.org/abs/2603.03338 in that it indicated (1) RSI is taken seriously by industry researchers, (2) the risk is seen as high by a large group, and (3) internal models are likely to be increasingly closed.
Severin Field 🦅
This past year, I interviewed 25 scientists at OpenAI, Anthropic, DeepMind, Meta, Princeton, Berkeley & Stanford about AI systems being used to automate the process of AI research, what is often called “recursive self-improvement” (RSI). I just published my updated takeaways…🧵