Interview, Fireside Chat
Gavin Miller: Adobe Research | Lex Fridman Podcast #23
- Adobe aims to evolve products to automate tedious tasks and shift creative roles from hands-on artisans to art directors or conceptual artists, leveraging AI to reason about intent for tasks like reflowing content across aspect ratios or languages.
- Plans include developing auto-buttons with content-based default settings, neural nets for zero-click selections of known categories, and "concept canvas" tools that assign spatial regions to keywords to establish user ownership.
- Adobe expects automatic tools to significantly reduce workflows, such as pulling selection masks from moving targets in seconds or performing sky replacement in one second, though the company notes that requiring 100% automation could delay market entry.
- The roadmap involves creating an ensemble of specialized models and a dispatcher to direct tasks based on confidence, initially focusing on specific workflows before branching out, with a goal of inferring invisible structures and 3D geometry from 2D inputs.
- Challenges remain in high-resolution robustness for generative methods filling missing holes, while future state-of-the-art research is required to advance current low-resolution stability.
- Adobe intends to balance privacy with potential for fine-grained data sharing, explore on-the-fly learning without permanent storage, and utilize deep learning for invisible 3D reconstruction in 2D workflows.
- Strategies for the future include an "assistant with a new skill" that avoids increasing visual complexity, smart defaults acting as educational "I'm feeling lucky" tools, and a central "Sensei" platform to accelerate the transition of ideas into products.
- The outlook covers expanded 3D design capabilities via augmented reality and real-time animation demand, requiring new UIs for fine-grained tasks and spontaneous movement using gross motor skills.
- Specific product initiatives include "Rush" for rapid professional video editing, distinct from "Premiere Pro," and the exploration of immersive environments where users can "dance with light" using real-time ray tracing and AI-modified shadows.
- Research into robotics involves building autonomous snake and spider robots with onboard desktop-level compute, vision accelerators, and biomimetic qualities, eventually utilizing artificial muscle materials to replace motors.
- Future AI agents are envisioned to have meaningful conversations with specific owners, explaining their reasoning, merging voice synthesis and vision to create personalities rather than striving for general human-level intelligence.
- Internal operations will prioritize researcher freedom based on impact rather than top-down directives, with interns exploring unproven ideas to generate publications and a portfolio for the next one to two years.
- Adobe anticipates the need for public education on AI capabilities to manage expectations regarding image authenticity, requiring multiple sets of evidence, and acknowledges that intelligent robots will feel familiar as users are already accustomed to similar virtual personalities.