Into the Impossible With Brian Keating AI Insider CEO: The Models Too Dangerous for You to See. Emad Mostaque - #550
45 snips
Apr 13, 2026 
Emad Mostaque, founder of Stability AI and creator of Stable Diffusion, speaks about the models that will never be released and an urgent 800-day transformation timeline. He explores why trillion-dollar labs lock away frontier models. He discusses diffusion, world models from video to 3D, personalized persuasive agents, and how AI reshapes organizations, religion, and human roles.
AI Snips
Chapters
Books
Transcript
Episode notes
Diffusion Models Learned Physics And Enabled 3D World Models
- Diffusion-based video models learned physical dynamics (cups dropping) and enabled conversion from video to 3D world models that run on consumer GPUs.
- Emad links these world models to early embodied cognition and future real-time 3D exploration tools like Genie.
RLHF Can Kill Creativity By Rewarding Safe Traces
- RLHF and reward shaping can reduce creativity by reinforcing safe, common behaviors, turning diverse reasoning into narrow, accountant-like outputs.
- Emad notes RLHF rewards activities (e.g., calculations) that make users happy but kill out-of-distribution creativity.
Audit Your Digital Footprint Before Agents Build Your Double
- Prepare for personal digital doubles by consolidating and auditing your digital traces before agents ingest them.
- Emad predicts agents will read emails, calls, and create tax-deductible, cheaper digital proxies that can replace human labor.
