Hugging Face's Clem Delangue on Open Source AI and the LLM Bubble | MTS Live

Episode Summary

Executive Summary: Clem Delangue argues that open source AI is essential for innovation, competition, and safety, warning against over-restricting model access. He contrasts a closing U.S. ecosystem with China’s strong open-source momentum, defends releasing capable models publicly, and highlights robotics as the next major frontier where openness could accelerate adoption and global competition.

Main Topics: Open source as the engine of AI progress (Priority: 5/5): Delangue frames open source as foundational to modern AI, citing the transformer architecture and broader internet infrastructure as examples of publicly shared technology that drove rapid innovation. U.S. vs. China open-source AI ecosystems (Priority: 5/5): He contrasts a recent U.S. shift toward closed APIs with China’s emergence as the leading contributor to open-source AI models and tools, noting growing startup and academic reliance on Chinese models. The ‘LLM bubble’ debate (Priority: 4/5): Delangue suggests the risk of overinvestment is highest in LLMs sold via closed APIs, pointing to expensive data-center buildouts, uncertain margins, and questionable long-term moats. Safety, regulation, and model release (Priority: 5/5): He argues that restrictions on model access are often overblown and that openness can improve safety by enabling defenders, while bad actors should be targeted through law and enforcement. Geopolitics and AI cooperation (Priority: 3/5): He hopes U.S.-China discussions will foster transparency and openness around AI, including collaboration on distillation and broader access to the technology. Robotics as the next frontier (Priority: 4/5): Delangue presents robotics as AI moving into the physical world, where open source can unlock new use cases, apps, and competition—especially against China’s dominance in robotics. Why Hugging Face won over GitHub for AI (Priority: 4/5): He explains that AI artifacts differ from code in scale and infrastructure needs, and that Hugging Face’s platform is better suited for large models, datasets, and private/public hosting.

Key Arguments: Open source accelerated the AI revolution because foundational technologies like transformers were shared publicly and could be studied, modified, and improved. The U.S. has become more closed in frontier AI, while China has become the strongest open-source contributor and is now widely used by startups and academia. If there is an AI bubble, it is more likely concentrated in closed LLM APIs than in AI broadly because of heavy capital spending and uncertain economics. Safety concerns about releasing models are often exaggerated; historically, fears around GPT-2 and other releases proved overblown. Open access can improve security because defenders need the same capabilities as attackers to build protections and respond effectively. The correct regulatory approach is to outlaw harmful actions like hacking, not to withhold general-purpose capabilities from everyone. Transparency and openness in international AI agreements could help more people access technology and enable cross-border collaboration. Robotics is a critical next step for AI because it creates new interaction modes and practical use cases beyond screens and phones. Hugging Face’s success comes from handling AI-specific storage and distribution needs at scale, not from simply being a code repository like GitHub.

Data Points: Hugging Face robot shipments: almost 10,000 - Delangue says the company has shipped nearly 10,000 Richie Minis worldwide. Robotics app ecosystem: over 300 apps - He notes that more than 300 apps have already been created for the Richie Mini. Platform data added: 2 petabytes in one week - He contrasts AI artifact scale with code hosting, saying Hugging Face added two petabytes of data in the last week alone. Movie-equivalent data scale: 500,000 two-hour movies - He uses this analogy to illustrate the volume of data uploaded to Hugging Face in a single week. GPT-2 timing reference: 6-7 years ago - He recalls that concerns about releasing GPT-2 openly were already prominent six to seven years earlier.

Pivotal Quotes: "The way you want to control it is untie everyone and then regulate or fight the bad actors." — Clem Delangue: On why AI capabilities should remain broadly accessible rather than restricted preemptively. "If there's a bubble, it's probably an LLM, but we'll see what happens in the next few months." — Clem Delangue: On where he sees the most speculative overinvestment in AI today. "It's kind of like the story of technology... you don't want to tie down everybody's hands because it's too dangerous." — Clem Delangue: His analogy for why restricting general-purpose technologies can slow progress and create new risks.

Implications: Listeners should expect continued tension between openness and control in AI, but Delangue argues the long-term winners will be platforms that maximize access, safety tooling, and global collaboration—especially as AI moves into robotics and the physical world.