OpenAI co-founder John Schulman and AI researchers Charlie O’Neill and Beren Millidge said recursive self-improvement may not trigger a near-term “intelligence explosion,” citing technical bottlenecks in generalization, continual learning and sample efficiency under the current Transformer and reinforcement learning paradigm. The researchers said large-model progress is shifting from pure compute scaling toward data efficiency and sparse or modular architectures. They identified reinforcement learning, distillation and continual learning from real-world deployment data as key areas that could help small and medium-sized labs compete with top-tier model developers. The interview projected AI could deliver a 10x productivity boost in specific knowledge-work scenarios over the next 1–2 years, while general intelligence exceeding human experts across all computer-completable tasks may still take 3–5 years. Setting alignment goals was described as one of the core human responsibilities likely to persist the longest.