Researchers Predict AI Could Surpass Top Human Experts in 3-4 Years

In a recent podcast episode, John Schulman, co-founder of OpenAI and chief scientist at Thinking Machines, joined Zyphra CTO Beren Millidge and Baseten's head of model training Charlie O'Neill to discuss timelines for recursive self-improvement (RSI) in AI. Schulman predicted that AI could surpass top human experts in all cognitive work performable on computers within three to four years.

The three researchers broadly agreed that a 10x productivity boost for AI researchers within two years is plausible, though O'Neill estimated a longer window of five to ten years, citing his own ability to absorb information and make optimal Bayesian decisions as a bottleneck. The conversation also highlighted a core challenge on the path to RSI: while AI systems are approaching human-level performance on "what to do," they still struggle with the higher-level capability of "deciding what to do."

On continual learning, they noted that both supervised fine-tuning and online policy distillation suffer from catastrophic forgetting after hundreds of micro-updates, causing models to lose earlier general abilities. Reinforcement learning preserves general capabilities better but injects limited new knowledge. They predicted model release cycles will accelerate from every three months to weekly, daily, and eventually hourly, by which point continual learning may truly be solved.

Schulman also expressed concern about diversity loss after RL training, with models repeating similar narrative themes and open-weight models increasingly resembling Claude due to widespread distillation, potentially leading to a "monoculture." The researchers also offered individual predictions on the arrival of all-around white-collar remote workers and the timeline for fully automated AI research.