Building the Automated AGI Lab: Core Automation's Jerry Tworek and Rohan Anil

Current AI progress relies heavily on scaling transformer architectures and reinforcement learning, yet these models remain limited by their inability to learn continuously from real-world data after deployment. Transformers suffer from catastrophic forgetting during fine-tuning and lack the computational depth required for complex, long-horizon tasks. To overcome these bottlenecks, research must shift toward architectures that enable test-time learning and end-to-end optimization of both pre-training and reinforcement learning. Jerry Tworek and Rohan Anil, founders of Core Automation, argue that the industry’s focus on short-term release cycles hinders the development of these more expressive, efficient systems. By automating the research process—including kernel generation and experimental iteration—they aim to build models capable of self-improvement, ultimately moving beyond the current paradigm where human intervention is required to update and adapt AI systems to evolving real-world environments.