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

In a three-person conversation, John Schulman (chief scientist at Thinking Machines), Beren Millidge (CTO of Zyphra), and Charlie O’Neill (head of model training at Baseten) examine why recursive self-improvement may not yield superintelligent AI by 2036, citing the lack of a true generalization breakthrough despite strong benchmark performance. Millidge likens the situation to Moravec’s paradox, arguing that excelling at narrow tasks does not automatically translate into transformative intelligence.