What's really open about open-weight AI?
Summary
The VergeCast hosts David Pierce and Jake Kastarnakis welcome Robert Hart to dissect the complexities of open-weight AI models and recent safety breaches. They analyze how the OpenAI Hugging Face hack exposed vulnerabilities in closed systems, sparking debates on regulation, China’s strategic advantage, and the industry’s inadequate self-policing. The episode concludes with news on Microsoft’s Xbox emulation plans and Apple’s brief Telegram removal.
Topics discussed
Intro: Microsoft AI support, Telegram ban, and Find My tags
Sponsors: Odoo and Indeed
Explainer: What are open weight AI models?
Benefits of open weights for developers and privacy
Geopolitics: US vs China approaches to AI openness
The Hugging Face hack and AI safety implications
Anthropic's stance and the debate over closed models
Critique of Anthropic's public relations strategy
Sponsors: Framer, Shopify, Odoo, and Indeed
Regulation: Will the industry self-police or need laws?
The 'Sandbox' problem: Testing models in isolation
Transparency and government review of AI models
Conclusion: The future of the open vs closed AI debate
Outro, listener hotline, and subscription info
Listen ad-free on Castria