AI researchers debate how close we are to recursive self-improvement

In the podcast episode featuring AI researchers John Schulman, Beren Millidge, and Charlie O’Neill, they discuss the possibility of achieving recursive self-improvement in artificial intelligence. Schulman agrees with Millidge's point that humans have advantages over models in various aspects, and each new model release brings temporary breakthroughs but often quickly becomes less impressive due to the model's limitations and lack of self-checking capabilities.

This cycle may continue indefinitely, with the model constantly lagging behind when used for actual research and engineering tasks. O’Neill adds that the gap between current AI systems and the ideal "learner you could have on a chip" is a significant factor in whether we will witness fast takeoff and recursive self-improvement. He believes that if we continue along the current trajectory of scaling up transformer models and RL, we might eventually hit this optimum point, leading to rapid advancements in AI capabilities.