Hugging Face's Clem Delangue: Don't Compare Engines to Cars

With a bit of tweaking, tools like ML InTern can make even a basic coding agent quite useful for AI development. By slightly adjusting the Hugging Face Hub integration and execution structure, it became capable of fine-tuning small models, generating datasets, and even converting model formats — and it passed a researcher interview test in just 30 minutes.

Clem Delangue believes the number of AI builders could grow from the current few hundred thousand to a few million, up to tens of millions, and potentially as many as 100 million. As a result, people would be less tied down by closed APIs and vendor lock-in, and less exposed to price hikes, model deprecations, and undisclosed changes.

• Interest expands into fields like biology, chemistry, healthcare, and climate.
• Solving real problems will take up a bigger share compared to simple video-AI slop.
• Hugging Face believes that as agents become users that pull models, use datasets, and read documentation, its user base could exceed the number of human users by the end of 2026.

His core argument is that simply comparing open models to closed APIs is meaningless. Behind an API, there are tools, routing, and multiple models working together, so what matters isn't whether 'open is falling behind,' but which system fits which task at what cost.

Reachy Mini is a case that illustrates this perspective. This open-source desktop robot, which has sold nearly 10,000 units, showed that when users spend 3 hours assembling it themselves and start building apps, perceptions of AI robots change quickly.

On the security debate, he pushes back against the claim that open source is more dangerous. He argues that the core of cybersecurity is making defenders stronger than attackers, and that open-source repositories get patched faster than closed systems. He also points out that GPT-2 was once said to be too dangerous to release, but the outcome turned out to be closer to overstatement.

He believes a company choosing not to open-source something can be a legitimate business strategy, but dressing up that reason as 'safety' is dishonest. Fear-based marketing does actually sell, and he notes that right after the Project Glasswing announcement, related companies sent commercial contract proposals. Even releasing just some papers, datasets, or small models can bring benefits in hiring, credibility, and visibility, with Mistral and Cohere cited as examples of this. At the same time, he warns that lobbying to restrict open-source AI is intensifying again in DC.