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

The current Transformer-based AI architecture has reached a performance plateau, necessitating a transition toward systems capable of continuous, test-time learning. While Transformers effectively distill internet-scale knowledge through pre-training, they lack the flexibility to adapt to evolving real-world tasks without constant, lab-based retraining. Former OpenAI and Google Brain researchers Jerry Tworek and Rohan Anil argue that future advancements depend on integrating pre-training with reinforcement learning and developing architectures that learn autonomously from user interactions. Their new venture, Core Automation, focuses on building an "autonomous lab" that accelerates research by automating complex kernel generation and optimizing end-to-end training processes. By shifting from static model scaling to dynamic, self-improving systems, they aim to overcome the computational inefficiencies and rigid development cycles that currently limit the practical application of frontier AI.