No Priors Ep. 3 | With Stability AI’s Emad Mostaque

Mostaque argues that governments, academia, and coordinated open efforts will have more compute and broader incentives than any single private company, making core base models an open infrastructure layer while businesses compete on instruction-tuning, fine-tuning, and applications.

He highlights work like Chinchilla, FLAN, and Med-PaLM to show that better data, instruction-tuning, and specialized fine-tuning can make smaller models highly capable, and predicts a future where many interoperating models outperform a single gargantuan system.

From Stable Diffusion to video and audio models, studios are saving millions, creators gain powerful tools, and visual communication becomes accessible to non-experts—transforming everything from film production to everyday meme-making and art therapy.

By building AI-first systems in places with minimal legacy infrastructure, Stability and partners are already demonstrating rapid gains (e.g., tablet-based literacy in refugee camps), which could outpace Western systems and later feed back improvements.

Projects like OpenFold, DNA diffusion, BioLM, and planned open Med-PaLM–style models aim to standardize and democratize powerful tools for protein folding, drug discovery, and medical reasoning, potentially bringing top-tier care insights to anyone with a phone.