Open Source AI? More Transparency, Please

Jockeying for position among the leading generative AI large language models (LLMs) has amplified their differences. Training models and code access are the source of some of the biggest disagreements. Should code for generative and other forms of AI be open or proprietary, protected under copyright, trade secret or even patent?

There is a lot riding financially on the outcome, and there are good arguments for and against.

ChatGPT’s parent, OpenAI, believes that closed code is safer and is more likely to thwart bad actors and encourage reliable development. (OpenAI’s product, ChatGPT, is the least open of the LLMs). Others believe that models, such as the open one suggested by Meta, are a more efficient and inclusive way to go.

AI2

Enter the Allen Institute for AI, the impressive nonprofit funded at least in part by Microsoft founder Paul Allen’s fortune. The Allen Institute for AI (AI2) believes even Meta’s proposed model does not go far enough, and sharing code leads to better outcomes for everyone.

AI2 and its sister nonprofit, the Allen Institute, do important work. It is not clear, however, where AI2’s $103 million 2022 revenue, almost all contributions, came from. Possibly from some of those Microsoft shares Paul Allen’s heirs’ control. In the 1980s, Microsoft co-founder Allen owned 25% of the company. Three years before his untimely death in 2017 from the effects of non-Hodgkin lymphoma, in January 2014, he still held 100 million shares of Microsoft, or approximately 5% of the company. At the current stock price of about $421 that is a value of about $42 billion. Bill Gates owns approximately 103 million shares for a current value of about $43 billion.

Former CEO and one-time Bill Gates’ assistant, Steve Ballmer, owns the most shares of Microsoft, the second most valuable company, with a market value over $3 trillion at this writing. His stake is currently about 4%, representing 90% of his $157 billion net worth. This has made him the seventh richest person in the world.

When he was alive, Allen made many savvy and successful investments, including the Seattle Seahawks (NFL), Portland Trail Blazers (NBA), Dreamworks SKG (movie studio), and more than two dozen other technology, media and content companies. It appears that his stake in MSFT was still his primary wealth engine. He also held 43 patents throughout his life.

Transparency Rules

Why am I bothering with all of this? It is because of a fundamental difference: AI2 is all fighting for open-source AI and ChatGPT, the current LLM leader, is definitively about closed code. Open AI also accepted a $13 billion Microsoft investment. Allen’s primary heir, his sister, Jody Allen, needs to be a bit more transparent about her ownership of Microsoft shares and how much of a part they play in funding AI2. There is a dramatic, potentially AI-defining difference in approach.

With all the double speak and positioning in this lucrative and possibly life-altering space, transparency is at a premium – and not just for data collection.

Allen left the MSFT board in 2000, citing health reasons. Had the eclectic investor and inventor broken with Microsoft’s proprietary approach that emphasized copyright and other IP rights, and licensing? Can you be a leader in truly open OS for AI and still own billions of dollars in Microsoft shares? Sure, why not? But it does send a rather mixed message and speaks to the need for more clarity in this still quite murky IP environment.

On May 11, 2023, AI2 announced they were developing OLMo, an open language model aiming to match the performance of other state-of-the-art language models. In February 2024, it was open-sourced, including code, model weights with intermediate snapshots and logs, and contents of their Dolma training dataset, making it the most open state-of-the-art model available, according to AI2’s website.

The Allen Institute, separate from AI2, is a non-profit, bioscience research institute located in Seattle. It was founded by billionaire philanthropist Allen in 2003. The Allen Institute conducts large-scale basic science research studying the brain, cells and immune system in effort to accelerate science and disease research. The organization practices what it calls open science and makes all their data and resources publicly available for researchers to access.

The questions are simple:

• How much Microsoft stock do Allen’s heirs, the Allen Institute, family foundation, et al., own? Have they been used to fund AI2?
• If the Allen heirs own as much as $42 billion in appreciated Microsoft shares, might these interests be seen as at odds with the company’s substantial position in OpenAI ($13 billion at last count), which sells the most used closed code LLM, ChatGPT?

It should not be confusing given the level of importance. Paul Allen’s sister, Jody, oversees his estate. She is not personally listed as among the top holders of MSFT. It is not clear if she sold some, most, or all the shares, or if they are held elsewhere. To be fair, perhaps for the Allen heirs, it is not just about profiting from the MSFT shares but using them to positively influence the company and encourage what is in their view a healthier world through more accessible AI models. Not everyone will see it that way.

Haves and Have Nots

In a perfect world it would be ok to own massive amounts of shares of the tech giant leading the strategy you oppose. It’s not a perfect world. That said, to compete in AI infrastructure, including training and code, businesses need serious investment capital. The software, NVIDIA chips and data mining are incredibly expensive. They are areas that China’s government and businesses are already investing in whether U.S. companies are prepared to or not. Bigger may indeed be better.

Early in the AI battle, the terrain is starting to resemble that of the early days of Internet; the space already seems to be shaking out into the well-capitalized haves and nimbler have nots. That does not mean that important, healthy development can’t be generated by different approaches. AI2 and Allen Institute do great work. It would not be wrong, however, to be more forthcoming about how the wealth of their benefactors is applied to the development of open AI. Transparency is not overrated. Even the hint of a shadow can put good intentions in an unflattering light.

Perhaps the Allen heirs are thumbing their collective nose at the company he co-founded, despite being among its largest and wealthiest shareholders. That would be the rock-n-roll thing to do; and we know that Paul loved rock. (He founded the Museum of Pop Culture in Seattle and was a huge fan of Jimi Hendrix). In some ways, I would like to believe that. Or is the AI2 and the Allen Institute simply encouraging MSFT/ OpenAI and others to pivot by saying “we care about people and the healthy development of AI – not just profits”?

In the debate between proprietary or closed GenAI code and a truly open-source model there is a question of security. Do open models make nefarious behavior easier, especially on the part of U.S. adversaries? There are differences of opinion here. Some believe that LLMs and data collection are so fast moving that the only way to stop bad actors is to let things evolve and stay a few steps ahead of them. Others believe that openness benefits not only bad actors but some tech giants who can afford to build proprietary products and services around standards, and who are adept at gobbling market share.

In the 1990s, lucrative parts of the Internet, including search and ecommerce, were carved up quickly and dramatically by a few players. GenAI may evolve in a similar manner, which may turn out to be less of a mistake than the inscrutable facts of digital life.