Mira Murati: Former CTO of OpenAI

Mira Murati joins The Daily Show to demystify how generative AI systems like DALL·E create original images. She explains that while many assume AI “copies” existing artwork, the model instead learns patterns across vast datasets of images and descriptions. Through this pattern learning, it can generate images that have never existed before, using relationships between text and visual concepts to produce original creations.

The conversation shifts to concerns raised by artists and creative professionals. Trevor Noah cites an art competition where an AI-assisted piece won, prompting debates about what counts as “real” art. Mira acknowledges these tensions and notes that OpenAI views systems like GPT-3 and DALL·E as tools—extensions of human creativity rather than replacements. She draws a parallel to ancient myths about helpers with extraordinary abilities, emphasising that innovation has always required balancing power with awareness of vulnerabilities.

Trevor highlights DALL·E’s ability not only to generate images but to expand existing ones, as in its reimagining of The Girl with the Pearl Earring. Mira describes this as an expansion of imagination itself, removing the constraints of canvas and allowing users to explore new creative possibilities.

The discussion then turns to misinformation. Mira outlines OpenAI’s guardrails, including removing public figures from training datasets and applying filters that block prompts involving violence, hate, or policy violations. While she stresses the importance of bringing the public into conversations about AI’s capabilities, she admits that misinformation remains one of the most difficult challenges.

This wide-ranging discussion brings together two of the world’s leading AI thinkers—Stanford’s Dr. Fei-Fei Li and OpenAI CTO Mira Murati—to unpack what’s most exciting in AI today and how the field must evolve to remain safe, human-centered, and beneficial.

What Excites Them Right Now

Fei-Fei Li describes a major new effort in robotic learning, where her team is developing a benchmark of 1,000 real-world robotic tasks grounded in how humans actually spend time (from government surveys of time-use). Rather than “toy tasks,” these benchmarks reflect everyday human needs—cleaning toilets ranks high; opening Christmas presents does not. She says robotics today feels like early-stage ImageNet: a foundational moment that could reshape the field.

Mira Murati highlights the continued progress of the scaling paradigm—large neural networks + massive compute + vast datasets—and how this formula has unlocked leaps in linguistic intelligence and visual understanding. The long-term goal, she notes, is AI systems that hold shared, human-like concepts of the world across modalities.