AI summary by TwinMind
Using AI to Empower Creators fr Roblox Studio Head Stef Corazza
Sequoia Capital
54m 47s
Roblox Business Overview & Scale
- 79 million daily active users, several hundred million monthly active users
- $29 billion market cap, $3.5 billion run rate revenue
- $800 million paid out to creators annually
- 70 billion hours of gameplay per year
- $600 million operating profit
- 46 million daily active users over age 13
- 3.8 million daily active voice users
- 90,000 new experiences/games published daily
- 15 million experiences played annually
- Vertically integrated: own data centers, hardware, app distribution, creation tools (Roblox Studio)
- Revenue sharing model: only take revenue when creators make money, no upfront fees
Steph Kaza's Background & Journey
- Started at Stanford 20 years ago focusing on computer vision/ML for human motion measurement
- Founded Mixamo in 2008 (video-based motion capture/animation)
- Acquired by Adobe in 2015, spent 7 years building 3D portfolio
- Built Adobe Stager, Adobe Aero for AR
- Joined Roblox to focus on generative AI for gaming platform
Creator Demographics & Current Development Process
- Several million creators monthly
- Average creator age: mid-20s (older than player demographics)
- 60-70% focus on world building/art
- 30-40% actually code
- Coding barrier prevents many from adding interactivity
- Games range from natural disaster simulations to concerts, fashion design, racing
AI Assistant Capabilities & Impact
- Launched code assist March 2023
- Full assistant includes: code writing/completion/debugging, documentation, asset creation
- Material generator and texture generator for 3D objects
- 180% increase in code creation for assistant users
- 60% increase in material creation for material generator users
- 30% increase in game publishing for assistant users
- Higher retention rates than any other Studio features
- All AI tools provided free to creators
Data Strategy & Community Partnership
- Obtained permission from "overwhelming majority" of creators to use their data for AI training
- "One of the largest datasets in the world" and "most multimodal"
- Includes code, images, 3D assets, audio, video, interactivity data, usage analytics
- Data used to "teach AI game development" rather than individual tasks
- Community gets free AI assistant in exchange for data access
Assistant Actions & Advanced Features
- Large-scale data model modifications via natural language
- Example: Select 57,000 non-pine trees, change leaf color to yellow
- Highest-scoring feature release in Studio in 2 years
- Focuses on "dishwasher" tasks users don't want to do
- Combines code generation with Studio's command bar execution
- Full awareness of scene data model for contextual operations
Future Vision: Control vs Intent Capture
- Paradigm shift from fine-grained control to intent capture
- Traditional tools: 30 years of surfacing more control to users
- New approach: multimodal input (audio, gestures, sketches + text)
- 2 orders of magnitude faster creation velocity
- Challenge: maintaining fine control for professional creators
- Progressive disclosure for advanced users
- Interoperability with external tools (Blender, Photoshop)
In-Experience AI & Gameplay Evolution
- Moving AI from Studio creation to in-experience gameplay
- NPCs as first stepping stone (hooking characters to LLMs)
- Personalized, always-different games based on player history
- Constrained creativity model (e.g., boat building with AI-generated components)
- Hybrid models mixing NPCs with real players
- Potential for pure NPC experiences (AI-generated entertainment)
User-Generated Content & Avatar Creation
- Avatar creation democratized through AI
- Text/image input to generate avatars, clothing, accessories
- Removes 3D modeling skill barrier
- "More people have great ideas than can execute them"
- Blurring lines between users and creators (like music industry evolution)
3D Foundation Model & Technical Architecture
- Announced 3D foundational model (open source planned)
- Digital synthesis of scenes/worlds from multimodal input
- Goal: teach AI complete game development, not just individual tasks
- 3D data advantage: spatial/temporal coherence, handles occlusions
- Physics engine: traditional approximations more effective than neural networks
- Moderation as biggest challenge for AI-generated content
Neural Rendering Vision
- Belief that neural networks will revolutionize game visual styling
- Decouple asset creation from visual style
- Real-time game restyling via text descriptions/reference images
- Maintain physics consistency while changing visual appearance
- Potential for player-chosen rendering styles
- Compute-intensive currently but expected to run on high-end phones within 5 years
Data Processing & Training Challenges
- Challenge: converting raw data to trainable format rather than data collection
- Unsupervised learning to extract creator intent signals
- Using gameplay analytics to identify "fun" elements for AI training
- Teaching AI to create games that are engaging, not just functional
