ChatGPT, DeepSeek, Or Llama? Meta’s LeCun Says Open-Source Is The Key

The progression of the Chinese open-source AI model DeepSeek (V3, R1 Zero, and R1) has taken the AI world by storm. The debate between open-source and proprietary AI models had remained somewhat academic, philosophical, and even ideological. But with DeepSeek R1 hitting performance marks previously reserved for OpenAI o1 and other proprietary models, the debate became a documented study case highlighting the virtues of open-source AI.

Yann LeCun, Chief AI Scientist at Meta, weighed in on the subject in a now viral LinkedIn post:

“To people who see the performance of DeepSeek and think: ‘China is surpassing the US in AI.’ You are reading this wrong. The correct reading is: ‘Open source models are surpassing proprietary ones.’ DeepSeek has profited from open research and open source (e.g., PyTorch and Llama from Meta). They came up with new ideas and built them on top of other people’s work. Because their work is published and open source, everyone can profit from it. That is the power of open research and open source.”

LeCun advocates for the catalytic, transformative potential of open-source AI models, in full alignment with Meta’s decision to make Llama open.

DeepSeek’s Open-Source Advancements

DeepSeek, a Chinese AI startup, has garnered significant attention by releasing its R1 language model, which performs reasoning tasks at a level comparable to OpenAI’s proprietary o1 model. Notably, DeepSeek has fully open-sourced R1 under an MIT license, allowing free commercial and academic use. This approach contrasts with the costly subscription models offered by competitors like OpenAI.

DeepSeek’s R1 model employs a multi-stage training pipeline that integrates supervised fine-tuning (SFT) with reinforcement learning (RL) to develop advanced reasoning capabilities. This approach has garnered significant attention from U.S. researchers, highlighting China’s potential to rival Silicon Valley in AI advancements.

The Case for Open-Source AI

Proponents of open-source AI, like LeCun, argue that openness fosters collaboration, accelerates innovation, and democratizes access to cutting-edge technology. By sharing models and codebases, researchers and developers worldwide can build upon existing work, leading to rapid advancements and diverse applications. Meta’s AI division, under LeCun’s guidance, has embraced this philosophy by open-sourcing its most capable models, such as Llama-3. This strategy aims to harness collective expertise to drive AI forward.

Mark Zuckerberg made the same case, albeit in a more explicitly business-focused manner, emphasizing that making Llama open-source enabled Meta to foster mutually beneficial relationships with developers, thereby building a stronger business ecosystem. This approach improves Llama and provides Meta with otherwise unattainable clarity about the market's direction regarding LLMs

Challenges of Proprietary AI Models

In contrast, proprietary AI models are often developed in isolation, with restricted access to underlying architectures and data. While this approach can lead to significant breakthroughs, it may also result in duplicated efforts and slower dissemination of knowledge. Moreover, proprietary models can create barriers to entry for smaller organizations or researchers lacking substantial resources, potentially stifling innovation.

Balancing Openness and Security in AI

Despite the advantages of open-source AI, concerns about security, misuse, and ethical considerations persist. Open models can be exploited for malicious purposes, prompting discussions about responsible AI development and the need for frameworks to manage openness. A paper titled “Towards a Framework for Openness in Foundation Models” emphasizes the importance of nuanced approaches to openness, suggesting that a balance must be struck between accessibility and safeguarding against potential risks.

LeCun addresses the openness-security debate by advocating for an open AI research and development ecosystem—with appropriate safety measures in place. He argues that this approach will drive progress, ensuring that “good AI” (advanced AI used by ethical actors) stays ahead of “bad AI” (trailing AI exploited by malicious actors). In his view, this tradeoff is advantageous in the long run, as a proprietary, closed approach to AI would never fulfill its greatest potential: providing universal access to knowledge and enabling intelligent, natural, and intuitive interactions.