Frontier AI Models Fail at Automated Research | Qasim Wani

Recursive Self-Improvement (RSI) represents the ability of an AI agent to invent the next version of its own research environment, yet current frontier models fall significantly short of this goal. With average accuracy on the GDPX benchmark hovering around 24 percent, models struggle to prioritize experiments or identify high-quality data, often failing to distinguish between valuable research and reward-hacking behavior. While some argue that RSI will eventually overcome data bottlenecks, high-quality data remains the primary constraint for scaling AI capabilities. Qasim Wani, co-founder of Extrapolate, emphasizes that true progress requires moving beyond simple automation toward creating agents capable of managing the complex, iterative, and data-intensive workflows that define modern AI research labs. Achieving superintelligence depends on engineering systems that can reliably audit data and navigate the nuanced, non-linear challenges of model development.