Core Automation is an AI research laboratory founded by Jerry Tworek, featuring research scientists and engineers previously affiliated with leading AI organizations including OpenAI, Anthropic, and DeepMind. The lab focuses on developing advanced AI systems capable of automating AI development itself, often described as building “an AI to build AI.” This work targets meta-level AI development and the automation of machine learning engineering processes.¹⁾

Core Automation represents a research initiative centered on automating the AI development pipeline. Rather than focusing exclusively on applications of artificial intelligence in traditional domains, the lab's research direction emphasizes using AI systems to improve and accelerate the processes through which AI systems themselves are created, trained, and optimized. This approach addresses a critical bottleneck in AI development: the resource-intensive nature of training large language models and managing complex machine learning workflows.

The founding of Core Automation reflects broader industry trends toward increasing automation of machine learning operations (MLOps) and the systematic optimization of training procedures. By bringing together researchers from multiple leading organizations, the lab combines diverse perspectives on neural network training, model architecture design, and system optimization.

The lab's work on automating AI development encompasses several technical domains:

Meta-Learning and AutoML: Research into automated machine learning systems that can discover optimal architectures, hyperparameters, and training procedures without extensive manual engineering.

Training Optimization: Development of systems to automate decisions about data curation, model scaling, compute allocation, and training procedures that typically require significant human expertise and experimentation.

AI-Assisted Development Tools: Creation of AI systems that can assist in code generation, model debugging, and performance analysis for machine learning applications.

Scaling and Efficiency: Investigation of techniques to improve the efficiency of AI model training and deployment, reducing computational requirements while maintaining performance.

These research areas reflect the increasing complexity of modern AI systems and the need for more efficient development methodologies as model sizes and training costs continue to scale.

The laboratory draws talent from three major AI research organizations. OpenAI, Anthropic, and DeepMind have each made significant contributions to large language model development and AI safety research. The combination of researchers from these institutions brings expertise in various aspects of AI development, from fundamental research to large-scale model training and deployment.

Jerry Tworek, who founded the organization, brings experience from work on AI development practices and machine learning systems. The interdisciplinary composition of the team positions Core Automation to pursue ambitious research directions that require coordination across multiple subfields of artificial intelligence.

Core Automation's mission to automate AI development itself addresses one of the key scaling challenges in the field. As AI systems become more capable and computationally intensive, the ability to automate aspects of their development becomes increasingly valuable. Successful advances in this direction could accelerate the pace of AI research and reduce the resource barriers to developing sophisticated AI systems.

The lab's focus on meta-level AI problems reflects a growing recognition that optimization of the AI development process itself may be as important as optimization of individual AI systems. This approach potentially enables more efficient resource utilization across the AI research and development ecosystem.

• Jerry Tworek
• Devin: Autonomous AI Coding Agent
• System-Level AI / Ambient Intelligence
• Anthropic Fellows Program
• How the Automation Cliff Is Driving a Blue-Collar Renaissance