AI Models Will Learn at Test Time by Next Year | Rohan Anil

Rohan Anil, co-founder of Core Automation and former Anthropic researcher, addresses the critical need for algorithmic breakthroughs to overcome current limitations in AI training. The conversation centers on the concept of "compute islands," a strategy for aggregating distributed GPU resources to reach frontier-level performance without requiring massive, co-located data centers. This approach leverages asynchronous training and second-order methods to maximize compute efficiency. Beyond hardware, the discussion highlights the "jagged" nature of current models, which struggle to learn from mistakes in real-time. Developing architectures that enable test-time learning remains the primary objective for achieving more reliable, autonomous systems. As labs increasingly verticalize their operations—including custom chip design—the ability to efficiently utilize distributed compute will determine which organizations successfully push the boundaries of artificial intelligence.