Lex Fridman PodcastGavin Miller: Adobe Research | Lex Fridman Podcast #23
At a glance
WHAT IT’S REALLY ABOUT
Adobe Research Chief Explores AI’s Future in Creativity and Robots
- Gavin Miller, head of Adobe Research, discusses how AI is transforming creative tools to automate tedious tasks while preserving and expanding human creativity. He explains concrete projects in image editing, search, and style transfer, and how Adobe balances power-user control with new, assistive interfaces. Miller also talks about data, privacy, and learning from user workflows, plus his parallel life as a poet and roboticist using biomimetic robots to explore intelligence and personality. The conversation ranges from deep technical ideas like GANs and 3D understanding to philosophical questions about altered reality, authenticity, and the future of human–machine collaboration.
IDEAS WORTH REMEMBERING
5 ideasAI will increasingly handle tedious production work so humans can focus on ideas.
Adobe’s goal is to automate low-level tasks like selections, background removal, aspect-ratio variants, and layout adjustments, freeing creatives to spend more time on story, style, and concept rather than manual pixel work.
Smart defaults and assistive tools can make complex software more accessible without sacrificing expert control.
Features like AI-driven ‘auto’ settings, one-click object selection, and suggestion systems based on similar user workflows aim to speed up beginners while still allowing professionals to refine results to 100% quality.
Understanding real workflows and online tutorials is key to building better creative assistants.
Adobe Research analyzes sequences of user actions and thousands of tutorial videos to infer common next steps, with the ambition to surface context-aware suggestions or mini-lessons directly inside tools instead of relying on manual search.
Modern image tools are evolving from single operations to full workflow macros powered by AI.
Examples like Sky Replace and Concept Canvas bundle selection, search, compositing, color matching, and layout into near-instant compound actions, enabling rapid exploration of multiple high-quality design alternatives.
Generative models bring ‘visual common sense’ but are limited by data, resolution, and domain.
GAN-based fills can plausibly hallucinate missing structure, but today work best at lower resolutions and on familiar categories; robust product features will likely require ensembles of specialized models and self-assessment of confidence.
WORDS WORTH SAVING
5 quotesWe want to free the artist to focus on inspiration and less on perspiration.
— Gavin Miller
Shipping is a different type of peer review.
— Gavin Miller
If you add a feature to a GUI, you add visual complexity; if you add a skill to an assistant, you don’t have to make the interface more intimidating.
— Gavin Miller
Being in research is a license to be curious.
— Gavin Miller
I like creating the illusion of life—if you coordinate enough degrees of freedom, it starts to feel like a creature rather than a thing.
— Gavin Miller
High quality AI-generated summary created from speaker-labeled transcript.
