Why I still haven’t bought into true RSI

The author discusses their hesitance in fully embracing the idea of true recursive self-improvement (RSI) in artificial intelligence. They point out that while the culture surrounding AI has become increasingly anxious and frenetic, especially within organizations like OpenAI and Anthropic, this heightened awareness might not necessarily translate into tangible progress. The author outlines several reasons why they doubt the imminent arrival of RSI.

Firstly, they argue that automatable research is too narrow to achieve significant acceleration in progress, given the exponential costs associated with scaling laws. Secondly, they note that the diminishing returns of adding more AI agents in parallel are a reality, and that resource bottlenecks and politics play a major role in building strong LLMs and AI systems.

These factors, combined with the fact that the labs have not yet unveiled any genuinely scary, specific breakthroughs, leave the author with high levels of uncertainty about the timeline for RSI.

The author contrasts this perspective with two recent podcasts from Dwarkesh, featuring Noam Brown and John Schulman, wherein the speakers discuss the role of RL, distillation, scaling, inference-time compute, and the potential for AI to solve known problems. While the authors generally agree with the claims presented, they find it surprising that the speakers predict timelines for AI capabilities that seem relatively short, such as an "okay" version of AI in one year or full generality within three years.

The author's conclusion is that while they acknowledge the potential for significant progress in AI, they remain skeptical about the timeline for true RSI. They believe that the current cultural and practical limitations make it unlikely that we will see superintelligence within the next eight years, despite the anxieties and rapid pace of progress in the field.