AI, Open or Closed Source?
An Unconventional, Controversial, and Thought-Provoking View
In the rapidly evolving world of artificial intelligence (AI), the debate between open source and closed source has reached a critical juncture. With the rise of Large Language Models (LLMs) like GPT-3, the question of accessibility and control has become a central concern. While the potential risks of open source AI have led to calls for restrictions, this article proposes a balanced approach: the establishment of an international end user registry mechanism to monitor and ensure responsible use of open source AI. This solution aims to harness the benefits of open source while addressing security concerns, all the while emphasizing the continued value of open source for AI explainability.
Open source software has a long history, with the most famous example being the Linux operating system. In the AI domain, the open source movement has given rise to accessible and collaborative platforms. Companies like Hugging Face have become central to the open source AI community, providing pre-trained models and tools that enable researchers and developers to build upon existing work. Other open source LLMs, such as LLaMa2, have also contributed to the democratization of AI technology. While these efforts have created great value in our society, the complexity and potential impact of AI, particularly Large Language Models (LLMs), introduce new challenges and considerations that differ from traditional open source software like computer OS.
The growing concerns over security and misuse have led some major companies to shift towards closed source development for LLMs. OpenAI's decision to make GPT-4 closed source, as discussed in The Guardian, is a notable example. Sam Altman, CEO of OpenAI, has warned about the risks of artificial intelligence, particularly the potential for large-scale disinformation and offensive cyber-attacks.
Similarly, Demis Hassabis, the CEO of DeepMind, expressed his concerns in a Time interview. He stated, “We’re getting into an era where we have to start thinking about the freeloaders, or people who are reading but not contributing to that information base. And that includes nation state as well.” Hassabis declined to specify which states he was referring to but hinted that it was fairly obvious. He further suggested that the AI industry's culture of openly publishing its findings might soon need to come to an end.
While open source AI is often praised for promoting accessibility and collaboration, the implications of such accessibility are multifaceted. On one hand, it democratizes AI technology, allowing a broader range of individuals and organizations to harness its capabilities. On the other hand, improved accessibility can level the playing field for global rivals. Many of these entities are unregulated, malicious, or operate under authoritarian or even totalitarian regimes. The lack of global rules and regulations can lead to misuse on a grand scale.
A striking example is nation state's use of AI for racial profiling and surveillance of the Uighur minority, as reported by The New York Times. Such instances underscore the potential dangers of unchecked accessibility. The democratization of AI, without proper safeguards, can inadvertently empower entities with intentions that run counter to global norms of human rights and ethical conduct.
Further intensifying these concerns, recent news reveals that Alibaba Cloud is set to support Meta's LLaMa 2 AI model in nation state. Alibaba, a company known to have assisted the ruling party in developing the app Xuexi Qiangguo as a party member tool, is now facilitating the use of an open source AI model in a country with a history of AI misuse. This move raises questions about the broader implications of open source AI models being utilized by companies with close ties to authoritarian governments.
So, is there a place for open source AI if we acknowledge such risks? Certainly! Open source can be an excellent tool for XAI.
AI explainability, often referred to as XAI (Explainable Artificial Intelligence), is the practice of making AI models' decisions understandable to humans. As AI models, especially LLMs, become more complex, their decision-making processes can become opaque, even to their developers. This lack of transparency can lead to mistrust and potential misuse of the technology. XAI tools aim to bridge this gap, offering insights into how models arrive at their conclusions, ensuring that AI operates in a manner that is both understandable and justifiable.
Such XAI tools can be completely open source, allowing everyone, from experts to the general public, to understand and interpret the decisions made by AI. This open approach ensures that the accountability and transparency of AI systems are reviewed constantly. By making these tools open source, it democratizes the process of AI scrutiny, allowing for a broader range of perspectives and expertise to contribute to understanding and refining AI behavior.
The Frontier Model Forum, a collaborative effort among tech giants, should consider opening such tools to the public. This would allow not only AI companies but also the general public to contribute to such scrutiny. In this proposed model, while LLMs remain closed source under government guidelines for safety and security, open source XAI tools, developed and refined by the public, will ensure AI's accountability and transparency. This dual approach balances the need for security with the imperative for transparency and public trust.
Recently, President Joe Biden signed an executive order to restrict U.S. investments in nation state technology, specifically targeting sectors like AI and semiconductors. This move, aimed at blocking and regulating high-tech U.S.-based investments towards nation state, underscores the intensifying competition between the world's two biggest powers. The order seeks to prevent nation state from leveraging U.S. investments in its technology companies to upgrade its military capabilities.
However, if the U.S. administration does not place restrictions on open source AI in U.S. companies, which have been heavily invested in with U.S. venture capital funding and human resources, the effectiveness of such executive orders could be undermined. If nation state companies can freely access and utilize open source AI developed in the U.S., it raises the question: What is the true meaning and impact of this executive order? It's imperative that the U.S. government considers the broader implications of open source AI in the context of national security and global competition.
While the concerns surrounding open source AI are valid, an outright block might not be the most constructive solution. Instead, we can harness the benefits of open source while ensuring responsible use. One potential approach is the establishment of an international end user registry mechanism. This registry would monitor the deployment of open source AI models, ensuring they're used ethically and responsibly. Entities, be it nations or organizations, wishing to utilize certain powerful AI models would register their intent, providing transparency about their applications.
Such a system would require international collaboration, emphasizing not the restriction, but the responsible use of AI. It's a balance between fostering innovation and ensuring global security. Challenges, such as enforcement and potential unregistered use, exist. However, with global cooperation and a shared commitment to ethical AI use, these hurdles can be addressed. It's time for global leaders, tech giants, and the AI community to come together, championing both innovation and responsibility in the AI landscape.
Update Jan 19th, 2024: Mark Zuckerburg didn’t commit into either open source or closed source if Meta achieves AGI level in this interview. Therefore, his decision to open source Llama-2 model, was motivated by the fact that the model is inferior to others and cannot generate revenue to Meta anyways? So, it’s a noise not a signal?
Updated March 17, 2024: x.AI has released Grok-1 into the open-source domain, marking a significant milestone as the most extensive language model in the open-source community, boasting 314 billion parameters. However, benchmark tests to evaluate its performance accurately are still pending. Despite this advancement, I think this is still a loser’s game that Elon Musk is playing. There's possibility that Elon Musk may have had leaked insights into GPT-5, suggesting a considerable technological gap between Grok-1 and GPT-5. This perceived disparity might have influenced Musk's decision to initiate legal proceedings against OpenAI and Sam Altman, subsequently leading to his decision to release Grok-1 as open-source.
Update Aug 1st, 2024: White House says no need to restrict ‘open-source’ artificial intelligence — at least for now after open source commnity’s lobby efforts. It’s clearly a national security risk. However, the report keep a future restriction open by stating “current evidence is not sufficient” to warrant restrictions on AI models with “widely available weights.” I sincerely hope we don’t need a catastrophic event to “warrant restrictions” .
(Inspired by multiple online and offline discussions with friends, with the help of ChatGPT)