Generative AI is Here. Who Should Control It?

Episode Summary

Executive Summary: Hard Fork explores the rapid rise of generative AI through a playful AI-voice stunt and a long interview with Stability AI CEO Ahmad Mostak. The episode frames generative AI as a major creative and commercial breakthrough, while pressing on concerns about misuse, copyright, bias, deepfakes, regulation, and whether open-source AI can avoid the mistakes of social media and centralized tech.

Main Topics: Generative AI as a breakthrough (Priority: 5/5): The hosts explain how new large language and diffusion models moved AI from analysis and ranking toward generating novel text, images, video, audio, and code. AI in creative workflows (Priority: 5/5): Both the hosts and external examples show generative AI already being used for illustration, concept art, interior design, advertising, and coding as a practical productivity tool. Open source versus centralized control (Priority: 5/5): Mostak argues Stability AI’s open approach gives users more freedom, customization, and global access than closed systems like OpenAI’s, which he says are too restrictive and centralized. Misuse, safety, and regulation (Priority: 5/5): The conversation addresses deepfakes, violent imagery, non-consensual sexual content, and government concern, with tension over whether trust-and-safety should be built in or decentralized. Copyright and artist concerns (Priority: 4/5): The episode raises worries that models are trained on scraped internet data and can imitate living artists’ styles, potentially enabling low-cost appropriation and displacement. Business model and market positioning (Priority: 4/5): Mostak outlines how Stability AI plans to monetize through API usage and enterprise/custom model services while presenting itself as a neutral alternative to big tech.

Key Arguments: Generative AI is different from older AI because it creates new content from prompts rather than only analyzing or ranking existing data. The rapid jump in capability came largely from scaling models and training data, producing emergent creative behavior. AI is already useful in real workflows, not just hype, because it saves time and expands what creators can produce. Open-source AI is more democratic and customizable than centralized models, especially for global/localized cultural needs. Safety should be addressed through community tools, filters, detection, and ongoing iteration rather than heavy preemptive locking-down. Human intention remains the key source of responsibility; the model is a tool, not an autonomous moral actor. The open release of models encourages innovation and broad adoption, while closed systems concentrate power in a few companies. Revenue can come from infrastructure, APIs, and bespoke model-building services rather than restricting access to the model itself.

Data Points: Funding round: $101 million - Stability AI’s seed round announced during the episode. Compute cluster size: 4,000 A100 GPUs / later 5,892 GPUs - Hosts discuss the company’s supercomputer scale; Mostak says it has grown beyond the earlier figure. Training dataset size: 100,000 gigabytes of images / 2 billion images - Mostak describes Stable Diffusion’s training data and compression into a smaller model file. Model file size: 2 gigabytes - Mostak says Stable Diffusion can run on an iPhone after being compressed into a 2GB file. Users of DALL·E: about 1.5 million users - Hosts cite usage as evidence of consumer traction for generative AI. Midjourney Discord membership: 3 million people - Used to show large organic adoption for image generation tools. GitHub Copilot adoption: more than 1 million programmers - Hosts note widespread code-completion usage. Company headcount: 100 employees - Mostak contrasts Stability AI’s size with large incumbents. Community size: 100,000 people - Mostak says the company’s strength comes from a large developer/community base. Deepfake detector prize: $200,000 - Mostak says Stability AI launched a prize for open-source detection tools. Open-source model downloads: 200,000 developers in just a few weeks - Mostak cites rapid adoption after Stable Diffusion release.

Pivotal Quotes: "The rest of this episode will be us. 100% human. 100% human. No AI voices at all." — Kevin Roose and Casey Newton: A disclaimer after the opening AI-voice bit, signaling the show’s focus on human judgment despite the topic. "The world has been creatively constipated, and we're going to let them poop rainbows." — Ahmad Mostak: Mostak’s vivid pitch for generative AI as a liberation tool for creativity and communication. "I believe this is one of the ultimate tools for freedom of expression." — Ahmad Mostak: His core justification for open sourcing and resisting centralized restrictions on generative models.

Implications: Generative AI is moving from novelty to infrastructure, reshaping creative work and platform power. The big question for users and regulators is whether open access will produce broad empowerment or accelerate harms faster than society can manage them.