Using AI to Empower Creators fr Roblox Studio Head Stef Corazza

Roblox generative AI head Stef Corazza previews a 3D foundation model, reveals 180% code-productivity lift from AI assistant, and argues 3D training data is necessary to fix video generation’s occlusion problem.

• Roblox creators using the AI assistant write 180% more code, create 60% more materials, and publish 30% more experiences than non-users.
• Roblox is building a 3D foundation model for scene synthesis from multimodal input; plans to open-source it.
• The platform pays out $800M/year to creators, runs $600M operating profit, and publishes ~90,000 new experiences daily.
• Overwhelming majority of creators consented to data use for AI training, giving Roblox what Corazza calls the world’s largest multimodal dataset (code, 3D assets, audio, video, analytics).
• Corazza argues 2D video data hits a hard wall for temporal consistency due to occlusions; 3D training data is the only real fix.
• Neural rendering (generative restyling pass over existing geometry) will be standard in game pipelines within 5 years, per Corazza.
• Moderation, not latency or compute, is the hardest problem for allowing in-experience AI-generated content on a platform with 79M daily users.
• Neural networks for physics simulation are not yet proven over classical approximations, at least for fully digital environments like Roblox.

2025-02-04 · Watch on YouTube