TL;DR:
- Meta’s Llama 2 release, though seemingly open-source, has limitations per Open Source Initiative (OSI) standards.
- Joelle Pineau, Meta VP for AI Research, acknowledges these limits as a necessary business balance.
- PyTorch, Meta’s open-source initiative, sets an example for generative AI models.
- Release decisions depend on code safety, prioritizing smaller groups for unproven work.
- Meta aims to foster collaboration in the open-source AI community.
- Meta’s approach contrasts with other big AI firms like OpenAI and Google.
- Smaller developers embrace open source and compete effectively in the commercial sector.
- Existing software licenses are ill-equipped for AI models that rely heavily on external data.
- Industry discussions highlight the limitations of open-source licenses for commercial AI models.
- OSI seeks to redefine licenses for AI models while preserving open-source principles.
Main AI News:
In the ever-evolving landscape of open-source software, Meta has stirred quite a debate with its recent release of Llama 2, a substantial language model. While Meta’s move to make Llama 2 accessible for free has garnered attention, it comes with certain limitations that leave some in the open-source community with reservations.
The crux of the issue lies in Meta’s licensing approach, which falls short of meeting all the criteria set forth by the Open Source Initiative (OSI). According to OSI’s Open Source Definition, true open-source software should encompass free redistribution, access to the source code, allowance for modifications, and the absence of ties to specific products. Meta’s restrictions include imposing a license fee for developers catering to an audience exceeding 700 million daily users and prohibiting other models from training on Llama. Researchers from Radboud University in the Netherlands went so far as to claim that Meta’s classification of Llama 2 as open source is “misleading,” and social media has echoed this sentiment, questioning Meta’s stance.
Joelle Pineau, Vice President for AI Research at Meta and head of the Fundamental AI Research (FAIR) center, acknowledges these limitations. However, she argues that striking a balance between sharing information and safeguarding Meta’s business interests is imperative. Pineau asserts that Meta’s limited approach to openness has transformed their research methodology internally, compelling them to prioritize safety and responsibility in their AI projects.
Notably, one of Meta’s most significant open-source endeavors is PyTorch, a machine learning coding language introduced to the open-source community in 2016. Pineau aspires to generate a similar level of enthusiasm for their generative AI models, especially since PyTorch has seen substantial improvements since going open source. Pineau emphasizes that the extent of their releases depends on factors such as the code’s safety in the hands of external developers. “Our decision to release research or code is contingent on the maturity of the work,” Pineau states. “When we are uncertain about potential harm or safety, we exercise caution by limiting access to a smaller group.“
FAIR is committed to ensuring that a diverse group of researchers can access their work for valuable feedback. This commitment extends to Meta’s approach to Llama 2’s release, emphasizing collaboration as the driving force behind innovation in generative AI.
Meta actively participates in industry groups like the Partnership on AI and MLCommons to contribute to the development of foundation model benchmarks and guidelines for secure model deployment. They advocate for collaborative efforts, believing that no single company should dictate the discourse on safe and responsible AI within the open-source community.
Meta’s approach to openness sets it apart in the realm of major AI companies. While OpenAI initially embraced a more open-sourced and open-research model, they shifted their stance, citing competitive and safety concerns. Google, too, has maintained a guarded approach to certain aspects of large language model development, occasionally sharing scientific papers but often withholding details.
In contrast, smaller developers like Stability AI and EleutherAI have thrived in the commercial space by championing open source. They frequently release new Large Language Models (LLMs) on platforms such as Hugging Face and GitHub. Falcon, an open-source LLM developed by the Technology Innovation Institute in Dubai, has gained popularity and is now rivaling both Llama 2 and GPT-4.
It is worth noting that most closed AI companies keep the details of their data gathering for model training datasets under wraps.
Pineau points out that current licensing schemes are ill-suited for software that incorporates extensive external data, a common practice in generative AI services. Most licenses, whether open-source or proprietary, provide limited liability to users and developers and minimal protection against copyright infringement claims. However, AI models like Llama 2 involve vast training data, potentially exposing users to more liability in cases of infringement. The existing software licenses do not adequately address this issue. “AI models differ from traditional software, as they entail greater risks. Therefore, we need to adapt our current user licenses to align better with AI models,” Pineau suggests, deferring legal considerations to experts in the field.
In the industry, discussions have arisen regarding the limitations of some open-source licenses for LLMs in the commercial sector. Some argue that the pure and genuine concept of open source is a philosophical debate, while developers prioritize other concerns.
Stefano Maffulli, Executive Director of OSI, acknowledges that current OSI-approved licenses may fall short in catering to the unique needs of AI models. OSI is actively exploring ways to collaborate with AI developers to ensure transparent, permissionless, yet secure access to models. “We must redefine licenses to address the real challenges posed by copyright and permissions in AI models, while preserving the core principles of the open-source community,” Maffulli states.
The OSI is also in the process of establishing a definition of open source specifically tailored to AI.
In the ongoing debate about whether Llama 2 qualifies as genuinely open source, it’s essential to recognize that this is just one facet of openness. A recent report from Stanford revealed that major companies with AI models often neglect discussions about potential risks and their accountability in case of failures. Acknowledging these risks and creating pathways for feedback should be standard practice for anyone involved in creating AI models, regardless of their open-source status.
Conclusion:
Meta’s nuanced approach to open source, with Llama 2 as a focal point, reflects a delicate balance between innovation and business responsibility. While this approach diverges from the strategies of some major AI players, it resonates with smaller developers, fostering competition and innovation in the commercial AI sector. However, the need to adapt licensing schemes to accommodate AI models signifies a growing challenge in the open-source landscape, calling for a reevaluation of industry standards to ensure transparent, secure, and collaborative AI development.
