When AI Starts Building AI
AI systems have traditionally been engineered by humans—from model architectures and training algorithms to serving systems and hardware optimization. This paradigm is beginning to change. As foundation models become capable software and systems engineers, they are increasingly able to automate the very infrastructure on which they run, opening the door to a self-improving feedback loop where AI builds better AI.
In this talk, I will describe that the future of AI development is an autonomous engineering stack powered by agents. I will first discuss recent evidence that frontier models are approaching expert-level systems engineering capabilities, including building high-performance inference engines from scratch. I will then present our work across the AI stack—from agent-designed serving systems for real-time multimodal applications, to new model architectures that improve efficiency and scaling, to training systems that dramatically accelerate agentic reinforcement learning. Together, these advances illustrate how agents can increasingly automate the design, optimization, and deployment of AI systems. While today’s agents are not yet fully capable of recursively improving themselves, they have already begun to transform how AI systems are built. The long-term opportunity is a self-evolving AI stack, where optimization becomes self-improvement and progress compounds with increasingly less human intervention.