
Building Large Design Models vs LLMs

From Figma to Code: The Rise of Design Engineers (And Why It Matters Now) - Honey Mittal
Topic 6 of 13
Published Oct 27, 2025 · Clip 32:13–48:57 (16:43)
Honey traces the evolution of Locofy's AI: from manual tagging, to CNN-based auto-suggestion of buttons and inputs, to graph neural networks for design optimization, to using OpenAI APIs pre-ChatGPT for naming layers. When ChatGPT launched, they evaluated LLMs but found their architecture ill-suited to understanding free-form Figma design structures, which are graph-like rather than language-like. They combine their own specialized models (transformers, YOLO, XGBoost) covering 80-85% of design-to-code with LLMs for code-level tasks. He advises experimentation rather than putting all eggs in the LLM basket, and argues LLMs will hit a scaling ceiling — they hit theirs long ago with design-to-code.