Interview, Fireside Chat
Gavin Miller: Adobe Research | Lex Fridman Podcast #23
- Adobe Research Mission: The lab focuses on automating tedious tasks and shifting creative workflows from "pixel space" to "idea space" using deep learning, aiming to increase "content velocity" and accessibility for users ranging from hobbyists to Hollywood professionals.
- Workflow Transformation: The organization is evolving from manual, low-level tool usage (e.g., selecting pixels, fixing aspect ratios) toward high-level intent-based actions where AI infers user goals and handles repetitive execution.
- Human-AI Collaboration Model: Products are designed to allow human intervention to bridge the gap from 99% to 100% robustness, prioritizing "smart defaults" and assistive suggestions over fully autonomous 100% automation to ensure professional-grade reliability without delaying market release.
- Learning Mechanisms: Research is analyzing thousands of hours of online tutorials to build "Adobe University" features that provide context-aware suggestions and inspiration based on observed user behavior patterns.
- Data Privacy Strategy: Adobe employs a "show, don't tell" approach where users are educated on immediate benefits (e.g., better recommendations) to encourage voluntary data sharing, with explicit opt-in options for detailed workflow data in exchange for tangible tool improvements.
- Project Sky Replacement: A single-action workflow that automates sky selection, matches stock content geometry, applies appropriate lighting recoloring to the foreground, and generates multiple mood variations instantly.
- Project Concept Canvas: A spatial search tool allowing users to define layout constraints (e.g., "dog on right, umbrella on left") rather than just text keywords, enabling users to feel ownership over generated designs.
- Generative Fill (Deep Fill/Project Scribbler): Uses GANs to infer invisible structures and fill missing image regions based on common sense knowledge, though current limitations restrict high-resolution stability, necessitating a future "ensemble of experts" approach for specific domains (e.g., houses vs. animals).
- Project Puppetron: A face-stylization technology that applies the texture of a painting to a moving video subject by reconstructing underlying 3D facial geometry to ensure motion consistency and avoid "uncanny valley" artifacts.
- AR/VR Design Philosophy: Adobe views VR as a tool for immersive transport and AR for contextual asset placement, with specific R&D focused on "spontaneous" 3D design in VR where gross motor movements replace precise mouse inputs for object manipulation.
- Tooling Strategy: The company maintains distinct product lines (e.g., Premiere Pro for precision, Rush for velocity) to avoid fragmenting the interface, offering radical simplification only on mobile or specific lightweight devices.
- Intern Program "The Funnel": Adobe utilizes a rotating intern program to explore high-risk, high-reward "fanciful" ideas without committing to long-term product investment, creating a pipeline for full-time researcher recruitment and academic publication.
- Research Culture: The lab operates on a model of "freedom of choice with impact-based rewards," avoiding top-down mandates to foster innovation while aligning projects with strategic company priorities.
- Future Tech Roadmap (2019+): Strategic goals include the "Sensei" platform for centralized, reusable neural models to accelerate feature deployment across all Adobe products, and integrating generative AI with real-time ray tracing for "dancing with light" simulations.
- Robotics & Biomimicry: Founder Gavin Miller's personal work on snake and spider robots converges with corporate research, focusing on adding onboard neural network compute for true autonomy and exploring artificial muscles to improve efficiency and biological mimicry.
- Philosophical Outlook: Miller draws parallels between modern digital curation (e.g., Instagram, deepfakes) and historical portrait painting, arguing that while the tools have evolved, the human desire to present "alternate realities" remains constant.
- Ethical Considerations: Adobe acknowledges the risks of AI-generated misinformation and deepfakes, emphasizing the need for public education on media literacy and the development of multiple evidence standards to verify authenticity.