Post-training to build better Roblox UIs (castform.com)

A recent collaboration between AI developers and Lemonade, a tool for creating Roblox games, has successfully post-trained a 295 billion parameter AI model, known as hy3, to enhance user interface (UI) generation specifically for Roblox. Historically, coding models have struggled with UI generation, particularly in specialized environments like Roblox, which uses the Luau programming language and specific APIs. The aim was to refine hy3 to produce visually consistent and appealing UI designs, meeting the unique stylistic demands of Roblox users while being a cost-effective solution compared to existing models like Grok 4.5. Key innovations in this project included developing a simulated Roblox Studio environment for testing and training the model, allowing it to create, edit, and navigate game elements as if operating within the actual application. The training process involved defining nuanced rewards based on visual fidelity and UI functionality, resulting in significant improvements over initial baseline performances. While the final model does not completely match the output quality of Grok 4.5, it offers a competitive alternative at a fraction of the operational cost, catering to the needs of Lemonade's growing user base. This initiative highlights the potential for post-training techniques to refine AI applications in niche domains, driving better user experiences and practical applications in game development.