“It would be a mistake for any country to try to slow down open source. A country that leads open source is a country that can lead AI in general” @ClementDelangue, co-founder & CEO @HuggingFace Watch the full interview on YouTube: https://youtube.com/watch?v=DfJV722V1WY&list=PLRRoCwK1ZTNCAZXXOswpIYQqzMgT4swsI…

Open Source is the Bedrock of AI Leadership: An Interview with Clem Delangue

TL;DR: Hugging Face co-founder Clem Delangue argues that it is a strategic error for any country to attempt to stifle open source. Open source not only accelerates the democratization of AI skills—empowering more people to build and thereby shaping public perception of AI—but is also key to maintaining technological leadership and preventing market monopolies. By lowering barriers and increasing transparency, the open-source ecosystem effectively counters fear-based marketing and ensures diversity and safety in AI development.

The ML Intern and Lowering the Barrier to Building AI

The interview begins by focusing on Hugging Face’s new ML Intern agent. Clem Delangue notes that default coding agents perform poorly when it comes to building AI. As Andrej Karpathy mentioned upon releasing Auto Research, early agents often fail to work effectively. However, by fine-tuning the model harness, the model itself, and its integration with tools like the Hugging Face Hub, ML Intern has made significant progress.

Currently, ML Intern can fine-tune small models, create datasets, and convert models into different formats. Excitingly, the team had the agent pass an interview test designed for researchers in under thirty minutes, with perfect performance. This demonstrates that if agents can lower the barrier to building AI, more people will be able to leverage open-source models and datasets, potentially even simplifying the historically complex process of deploying local models.

From Elite to Mass Adoption: A Surge in AI Builders

Clem predicts that the number of people with AI-building skills will explode from the current low hundreds of thousands or low millions to tens of millions, possibly even reaching 100 million. This means that in the future, nearly every software engineer may have the capability to optimize, train, and fine-tune models.

This shift has profound implications:

1. Reclaiming Control: Developers will no longer rely solely on closed-source APIs and third-party vendors. Closed-source providers can unilaterally change terms, raise prices arbitrarily, deprecate models, or make backend changes that degrade workload quality. Open source hands control back to the builders.
2. A Broader User Base: Unlike software engineering, which requires learning programming languages, AI is driven by datasets and text, appealing to a wider potential audience. Steve Yegge has pointed out that non-technical individuals will also enter the world of programming, a view Clem agrees with.
3. Solving Real-World Problems: Increased diversity among builders will bring broader perspectives. Currently, Silicon Valley leaders may focus more on entertainment applications, but a more diverse group of builders could drive practical AI applications in urgent fields such as biology, chemistry, medicine, and climate change, reducing what is known as “video AI slop.”

Changing Public Perception of AI: From Fear to Empowerment

Public perception of AI is currently predominantly negative, filled with fear or hatred. Clem believes that empowering the public to become builders is key to changing this narrative.

• Experience Over Preaching: Take Hugging Face’s physical product, Rich Minions, for example. While people might feel indifferent toward abstract “AI robots,” after assembling one (taking about 3 hours), building an application, and interacting with it, they often fall in love with the technology.
• Dispelling Marketing Hype: There is currently a lot of fear-based marketing in the market (such as campaigns related to Project Glasswing) designed to sell anxiety for profit. Allowing the public to build systems themselves helps them realize that AI is just a tool (Software 2.0 or 3.0), not a self-aware entity like the one in RoboCop. This sense of empowerment can counteract the impact of fear-based marketing.

Open Source and Security: Defenders vs. Attackers

Addressing concerns that “open-source models lead to weaponization or deepfakes,” Clem offers a rebuttal from a cybersecurity perspective:

1. Resilience Over Closure: The core of cybersecurity lies in giving defenders more power than attackers, ensuring the cost of attack exceeds the cost of defense. Open-source systems typically patch vulnerabilities much faster than proprietary systems. When proprietary systems are attacked behind closed doors, attackers may have weeks to exploit vulnerabilities before patches are deployed, which often lag behind.
2. Balancing Power Asymmetry: Closed source increases the asymmetry of power and capability, granting immense power to a few while leaving defenders unable to respond. Open source maintains balance, empowering defenders to fight back.
3. Lessons from History: Take GPT-2, for instance. Once considered too dangerous to release, it ultimately did not cause systemic issues today. The real risk lies in model leaks or restricting access to specific entities while denying it to the rest of the world, which only creates a false sense of security.

Harmonizing Business Models with Open Source

Regarding companies like ElevenLabs that choose not to open-source due to their business model, Clem expresses full understanding but opposes using “safety” as an excuse for opacity.

• Honest Communication: If a model is not open-sourced for commercial reasons, this should be stated candidly.
• Benefits of Partial Open Source: Companies can build credibility and visibility by publishing research papers, partial datasets, or small models, while keeping large models proprietary. Companies like Mistral and Cohere have proven that open-source strategies can lead to significant commercial success and attract top talent.
• Core Stance: There is nothing wrong with a company choosing not to open-source, but it should not mislead the public into believing this is done for safety reasons.

Policy Recommendations: Do Not Stifle Open Source

Clem emphasizes that the current global trend of stifling open source is a step backward.

• Source of Leadership: The United States’ leadership in AI is largely derived from its leadership in open source. For example, Google’s open-sourcing of the Transformers architecture (“Attention Is All You Need”) triggered widespread imitation and collaboration, laying the foundation for today’s technology.
• Strategic Risk: If the U.S. or other countries slow down open source, they will lose AI leadership within months or years.
• Preventing Monopolies: Stifling open source increases the concentration of power, capability, and revenue, leading to AI being monopolized by large tech companies like OpenAI and Anthropic. Imagine a world where only a few companies can develop software—it is terrifying. Open source fosters competition, imitation, and job creation, ensuring that the value created by AI is widely distributed rather than captured by a handful of companies.

Source: https://www.youtube.com/watch?v=DfJV722V1WY