Linux Foundation’s adoption of OMI could pave way for ethical LLMs, analysts say

The Linux Foundation’s move to take the Open Model Initiative (OMI) under its wing could pave the way for “more ethical” large language models (LLMs), analysts say.

“One of the core objectives for OMI and its induction into the Linux Foundation is to propagate an ethical use of data (text/images) to train generative AI models,” said Abhigyan Malik, practice director of data, analytics, and AI at the Everest Group.

However, Malik warned that the practice of training models with ethical data will become increasingly difficult, given the wider understanding of data protection and with popular sources changing their privacy and usage policies. 

Several proprietary LLM providers, such as OpenAI and Stability AI, are currently facing lawsuits that claim that these companies violated copyrights while training their models.

What is the Open Model Initiative?

The Open Model Initiative (OMI), which was founded in June by three startups — Invoke, Civitai, and Comfy Org, aims to bring together developers, researchers, and enterprises to collaborate on advancing open and permissively licensed AI-related model technologies.

Permissive licenses, according to the Linux Foundation, tend to make it easy for community members to participate and share contributions without downstream obligations.

“This particularly favors software segments that require the ability for software producers to distribute proprietary software based on the open source code base without revealing their changes,” the Linux Foundation explained in its guide for open source software.

The OMI’s core objective is to bring together deep expertise in model training and inferencing to develop models of equal or greater quality than proprietary models, such as LLMs from the stables of OpenAI, Google, and AWS, but free of restrictive licensing terms that limit the use of these models.

In order to achieve this, the OMI, which will be governed by a community-led steering committee, will establish a governance framework and working groups for collaborative community development.

The OMI will also conduct a survey to gather feedback on future model research and training from the open source community, the Linux Foundation said in a statement, adding that it will further create shared standards to enhance model interoperability and metadata practices.

Additionally, the OMI will develop a transparent data set for training, and create an alpha test model for targeted red teaming.

The ultimate goal of the initiative, according to the Linux Foundation, will be to release an alpha version of the model, with fine-tuning scripts, to the community by the end of the year.

Why is this of significance to enterprises?

The significance of this move for enterprises lies in the unavailability of source code and the license restrictions from LLM providers such as Meta, Mistral and Anthropic, who put caveats in the usage policies of their “open source” models.

Meta, for instance, according to Everest Group’s other AI practice leader Suseel Menon, does provide the rights to use Llama models royalty free without any license, but does not provide the source code.

“Meta also adds a clause: ‘If, on the Meta Llama 3, monthly active users of the products or services is greater than 700 million monthly active users, you must request a license from Meta.’ This clause, combined with the unavailability of the source code, raises the question if the term open source should apply to Llama’s family of models,” Menon explained.

In contrast, the OMI’s objective, according to analysts, is to create models that don’t present enterprises with caveats and are more freely accessible.

Will OMI stand before the might of Meta and larger LLM-providers?

The OMI’s objectives and vision received mixed reactions from analysts.

While Amalgam Insights’ chief analyst Hyoun Park believes that the OMI will lead to the development of more predictable and consistent standards for open source models, so that these models can potentially work with each other more easily, Everest Group’s Malik believes that the OMI may not be able to stand before the might of vendors such as Meta and Anthropic.

“Developing LLMs is highly compute intensive and has cost big tech giants and start-ups billions in capital expenditure to achieve the scale they currently have with their open-source and proprietary LLMs,” Malik said, adding that this could be a major challenge for community-based LLMs.

The AI practice leader also pointed out that previous attempts at a community-based LLM have not garnered much adoption, as models developed by larger entities tend to perform better on most metrics.

“A prime example for such an open LLM is BLOOM, that successfully created a community model but has not yet been able to create adoption due to inefficiencies and certain design choices (it was designed to not be a chat interface),” Malik explained.

However, Malik said that the OMI might be able to find appropriate niches within the content development space (2D/3D image generation, adaptation, visual design, editing, etc.) as it begins to build its models.

“These niches are aligned to various use cases (e.g., 3D image generation) or applications in the verticals (e.g., catalogue image generation/editing for retail) where its models may perform tasks effectively,” Malik said.

Malik’s theory may hold water, given Invoke is a generative AI platform for professional studios and Civitai is a generative AI hub for creators.

One of the other use cases for the OMI’s community LLMs is to see their use as small language models (SLMs), which can offer specific functionality at high effectiveness or functionality that is restricted to unique applications or use cases, analysts said. 

 Currently, the OMI’s GitHub page has three repositories, all under the Apache 2.0 license.