John Schulman, Beren Millidge, and Charlie O'Neill argue RSI faces hard limits: AI can 10x well-specified research but cannot think its way to new training paradigms.

Three researchers identify key blockers to recursive self-improvement: AI can accelerate research dramatically when objectives are well-specified, but cannot discover new paradigms because 'thinking alone can't generate new bits.' The likely ceiling is not a hard wall but a series of discontinuities — like the jump from pre-training to RL — each requiring a breakthrough current models may not be able to generate. John Schulman argues the last durable human role is defining what we actually want, i.e., alignment and objective specification.