We're training a Figma-fluent editing model (twitter.com)

A new model specifically designed for Figma editing workflows has been announced, showcasing the potential of specialized post-training techniques to enhance AI capabilities in non-verifiable design tasks. The model focuses on streamlining the iterative process commonly undertaken by designers, enabling quicker and more accurate edits in the Figma environment. Unlike traditional frontier models, the Figma agent utilizes supervised fine-tuning (SFT) and reinforcement learning (RL) to adapt to the nuances of design requests, allowing it to generate changes based on Figma's native file format. This approach emphasizes fast execution, which is crucial given the frequent, minor adjustments designers make throughout their day. The significance of this development lies in its ability to provide a competitive edge over existing models by enhancing both the speed and quality of design edits. With more than 250,000 training examples and the incorporation of RL, the Figma agent is positioned to better understand user intent and produce high-quality outputs. By combining internal evaluations and expert feedback, the model distinguishes itself in the crowded landscape of design AI tools, empowering designers to realize their creative visions more efficiently. This initiative also highlights a broader trend in the AI/ML community toward specialized applications that can outperform general-purpose models in specific domains.