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

Current transformer architectures face a significant bottleneck in scalability and adaptability, as they rely on static, lab-based training rather than continuous, real-world learning. While transformers have mastered pre-training and reinforcement learning, they struggle with catastrophic forgetting and data efficiency, limiting their utility for dynamic, complex tasks. Jerry Tworek and Rohan Anil, founders of Core Automation, argue that the next breakthrough requires a fundamental shift in architecture to enable models that learn at test time. By automating the research process—specifically through high-performance kernel generation and end-to-end optimization of the deep learning stack—the team seeks to replace the current, inefficient paradigm. Their mission focuses on creating systems that can autonomously improve themselves, ultimately aiming to achieve a level of intelligence that transcends the limitations of current, human-dependent AI development cycles.