Why Big Tech Is Making a Fuss About Open Source AI Models — It’s All About Ownership, Privacy, and Security

History is repeating itself, but this time it is about the total cost of ownership (TCO), reliability, accountability, privacy, and security of open-source artificial intelligence (AI) models.

Notable open-source AI models include Llama from Meta, Stable Diffusion from Stability AI, Eleuther AI’s Pythia large language model (LLM) suite and its smaller language model GPT-NeoX, Hugging Face’s BigScience Large Open-science Open-access Multilingual Language Model (BLOOM), and Databricks’ Dolly LLM.

However, there is a lot of confusion and disagreement about what exactly constitutes an open source AI model.

Here’s one reason: While Google, Meta, Microsoft, and Elon Musk’s xAI say they are promoting the use of open-source AI models, others, including OpenAI, Apple, and Nvidia, are seen as relying on closed-source models that keep their AI technologies proprietary for strategic gain.

However, OpenAI, which started its journey as an open-source company, is now closed. Google’s large Gemini model is also closed, but its smaller Gemma model is open. Even Musk’s Grok LLM does not qualify as a fully open AI model, according to the Open Source Initiative (OSI), the body that defines what open source means.

In addition, Apple, which is usually known for its proprietary ecosystem, now has its own OpenELM family of small language models, spanning 270 million to 3 billion parameters, that is open and tailored for mobile devices and computers. And Nvidia has also begun to open-source some of its graphics processing unit (GPU) drivers, which will benefit Linux developers.

New definition

Last week, the OSI said that for an AI system to be considered open source, it must be free to use for any purpose without permission. Users should also be able to study and research the AI ​​system’s components, modify it to change its output, and make the system available to others, whether in its original form or with modifications, for any purpose.

Mark Zuckerberg, founder and CEO of Meta, has been insistent that his company is committed to open-source AI. In a July 23 memo, he said that while many large tech companies were developing “closed” models, “open source is rapidly closing the gap.” Zuckerberg cited the example of how large tech companies developed proprietary versions of Unix, believing that closed systems were essential for advanced software. However, open-source Linux has gradually gained popularity due to its flexibility, affordability, and expanding capabilities.

Over time, Zuckerberg noted in his memo, Linux has surpassed Unix in security and functionality, becoming the industry standard for cloud and mobile operating systems. Zuckerberg said he believes AI will follow a similar path.

According to him, Llama 3 is already competitive with the most advanced models and leads in some areas. Llama, he adds, is already “a leader in openness, modifiability and cost-effectiveness.”

While this claim is unfounded, the fact is that Stanford University’s AI Index, released in April, reveals that organizations have made 149 foundation models available, of which 65.7% were open source, while in 2022, the figure was just 44.4% and in 2021, 33.3%.

Salesforce also recently released a new suite of open-source, large-scale, multi-modal AI models this month called xGen-MM (also known as BLIP-3).

“We make our models, carefully curated large-scale datasets, and our code base available for refinement to facilitate further progress in research on large multimodal models (LMMs),” the authors said in a paper published on arXiv.

India bull

Closer to home, Bengaluru-based startup Sarvam AI released in mid-August what it billed as India’s “first fundamental, open model” of a small Indian language. Called Sarvam2B, it’s a 2 billion-parameter model that was “trained from scratch” on an internal dataset of 4 trillion tokens and covers 10 Indian languages. Sarvam AI simultaneously released Shuka 1.0, an open-source AudioLM that is an audio extension of Llama 8B to support voice-to-text in Indian languages.

While these announcements were part of the launch of Sarvam AI’s Generative AI (GenAI) platform, which includes voice-enabled multilingual AI agents, an application programming interface (API) platform to make it easier for developers to use these models, and a GenAI workshop aimed at lawyers, the emphasis was on the platform and AI models being “open,” as opposed to “closed or proprietary.”

In June, Tech Mahindra announced a partnership with Dell Technologies under the LLM Indus project, which aims to “leverage AI-optimized technologies with an open ecosystem of partners…”

Similarly, AI4Bharat, a research lab at the Indian Institute of Technology Madras, is working with Bhashini to create datasets and ParamSiddhi of CDAC Pune to train models, with an emphasis on supporting “open-source tools and models.”

Challenges remain

Unlike proprietary models, which can be restrictive and expensive, open-source models are freely available for modification and integration. This flexibility allows companies to experiment with cutting-edge AI technologies without being locked into vendor-specific ecosystems. For example, companies like Tesla used open-source AI tools to build their self-driving technology, allowing them to rapidly iterate and improve.

Open source AI also fosters innovation by enabling collaboration within a global developer community. For startups and smaller companies with limited budgets, open source AI provides access to powerful tools that would otherwise be out of reach.

But open-source AI comes with its own set of challenges, particularly around total cost of ownership (TCO), security, and the need for skilled talent. In addition, open-source AI models, while highly configurable, cannot always meet the rigorous security standards required by enterprises, a point often made by big tech companies promoting closed-source AI.

In its report on “Dual-Use Foundation Models with Widely Available Model Weights,” released July 30, the Commerce Department’s National Telecommunications and Information Administration (NTIA) recommends that the U.S. government develop new capabilities to monitor potential threats, but refrain from immediately restricting the wide availability of open model weights in major AI systems. “Open weights” models in AI refer to models in which the trained parameters, or “weights,” are publicly available.

This transparency allows researchers and developers to examine, modify, or extend the internal structure of the model, while also allowing developers to extend and adapt previous work. This makes AI tools more accessible to small businesses, researchers, nonprofits, and individuals, according to the NTIA report.

However, as the Electronic Privacy Information Center (EPIC) noted in its March 27 comments to NTIA’s request for comments, while making model weights widely available “may promote more independent evaluation of AI systems and greater competition than closed systems,” closed AI systems are less vulnerable to adversarial attacks and allow for easier enforcement than open systems.

In its memo, EPIC called on NTIA, among other things, to “consider the nuances of the benefits, drawbacks, and regulatory hurdles that arise for AI models across the openness gradient—and how the benefits, risks, and effective oversight mechanisms change as models move along the gradient.”

However, this approach, while reasonable and balanced, is easier said than done.