Why I still haven’t bought into true RSI
The author discusses their perspective on recursive self-improvement (RSI) in AI, arguing that current trends suggest a "lossy self-improvement" rather than a rapid, unpredictable acceleration towards superintelligence. They note that frontier AI labs like OpenAI and Anthropic are using thousands of concurrent agents, which increases anxiety about AI progress and risks. However, the author believes that automatable research is too narrow, diminishing returns of AI agents are real, and resource bottlenecks limit progress. The article also references discussions on AI timelines from a podcast, where experts predict varying timelines for AI reaching human-level capabilities in specific tasks. The author concludes that the hardest exponential to overcome is peak intelligence, and while inference-time compute will scale, true RSI leading to dramatic acceleration is not yet evident, making increased discussion of extinction risk misplaced.
Original sources
Newsletter Interconnects
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.