Yesterday, accompanying Meta's release of Llama 3.1, Mark Zuckerberg shared a persuasive blog post championing open source AI and community-driven innovation. I loved it, and highly encourage you to read it. However, his narrative, while compelling, overlooks critical considerations that require a more nuanced examination.
This article directly rebuts several of Zuckerberg's key claims, highlighting significant gaps in his argument for open-sourcing frontier AI models. As always, my goal is to encourage more public discourse around critical AI topics like this.
On Meta...
The comparison between Unix/Linux and the current state of AI is interesting but misses a critical lesson for Meta. That is, who benefited from open-sourcing Linux? While Linux became the industry standard, the closed variants of Unix evolved into MacOS and iOS—platforms that are not only more profitable but also epitomize innovation and user experience.
Similarly Apple developed and open-sourced WebKit, only to see it adopted by rival Google who controls 65% of the market compared to Apple's 17% (50% of which is due to Safari being the default browser on iOS). There is no guarantee that the success of open source technology will lead to success for the creator. Meta’s open source AI initiative may not sustain its momentum if it fails to monetize effectively, especially when rivals benefit from the same technology without incurring the development costs.
The reality is that many open source projects today often rely on the benevolence and deep pockets of corporate sponsors. And for models like Llama, this is truer than ever. Llama is a donation from Meta, that we gladly accept. But, let us not kid ourselves, Zuckerberg's argument for open source rings hollow when Meta does not release the training data for Llama models. Of course, this is their choice—but truly embracing open source would mean sharing not just the models but also the datasets and methodologies. By withholding this information, Meta undermines its own argument for openness, suggesting that strategic advantages does in fact lie with proprietary IP.
On China...
I see several issues with this line of argument, primarily concerning security and resource allocation. Firstly, the truth is that only a handful of companies possess the resources to develop frontier models. Therefore, the onus of securing these advanced technologies really isn't on "most tech companies" but only a few entities. Surely, we can invest in keeping those secure.
In any case, if AI is indeed as dangerous as these companies suggest, we should be demanding the highest standards of operational security from those developing and deploying these technologies. The notion that we should make potentially dangerous technology freely available simply because some adversaries might steal it is a non-sequitur.
Additionally, why hasn't China done this with Gemini/Anthropic/Claude? Over the last year, China did not drastically acclerate its AI capabilities despite the availability of capable models from these companies. Instead, evidence suggests that China’s advancements have primarily leveraged open-source models like Llama, rather than pilfered proprietary technologies.
On Innovation...
Ben Thompson in his Stratechery Update endorsed Mark's vision:
Historical evidence clearly shows that innovation does not necessarily require openness. Apple and NVIDIA are prime examples of companies that have thrived by keeping key technologies proprietary. Apple, a paragon of innovation, has thrived by keeping its critical iOS technologies proprietary. In contrast, Google’s decision to open source Android has aided rivals like Huawei and Samsung in surpassing it. Similarly, NVIDIA, has not open-sourced its core technologies and yet continues to dominate the GPU industry. The blanket assertion that innovation necessitates complete openness is demonstrably false.
One of the advantages of open source is community contribution. However, AI development does not mirror traditional software development. While thousands of forks of models like Llama exist, there is no streamlined process for integrating these improvements back into the primary model. This decentralization dilutes the potential benefits of community-driven enhancement, limiting the actual progress achieved through open source.
On Security...
When it comes to the open vs. closed debate, we don't need to look further than the AI ecosystem itself. Despite the significant advancements represented by Llama 3.1, it only matches or slightly surpasses the year-old GPT-4. OpenAI’s impending release of GPT-5 is expected to once again widen this gap. Furthermore, the security argument for open source is left wanting. Is Llama 3.1 more secure? No. In fact, it was jailbroken even before it was released publicly.
Additionally, closed source models, like those from OpenAI, can implement real-time updates and additional protections when things go wrong. When, Gemini's image generation capabilities went off the rails, Google was able to pull it. In contrast, open source models, once released, cannot be retracted or modified to address emergent risks.
On Regulation...
So where does this all lead? Yesterday, I responded to Nicholas Thompson's monologue on the release of Llama 3.1, and highlighted the irony that the US bans chips from China while remaining silent (maybe clueless) about the release of open source models.
This discrepancy underscores the need for AI literacy and informed regulatory frameworks. While I don't think we are there yet, as these systems become more capable, regulatory bodies should be actively involved in assessing and approving any open source release of frontier AI models—ensuring that their deployment aligns with national security and public safety interests.
Open source AI obviously holds significant value for fostering innovation and democratizing technology. However, acknowledging its risks and implementing robust security measures are crucial. Open-sourcing frontier models without some sort of oversight seems irresponsible. To this end, thankfully we have seen a lot of open research being shared on this front.