SXSW Retrospective: Exploring Open Source AI with MoFo, Meta, Databricks, & EleutherAI

Listen to the full audio recording.

At the 2025 South by Southwest Conference (SXSW) in Austin, Texas, MoFo partner Justin Haan led a dynamic panel discussion exploring the legal, business, and policy challenges emerging from the rapid growth of open-source artificial intelligence. In light of the release of increasingly powerful models, the panel delved into key questions around the meaning of “open-source” AI, the business and legal considerations driving the release of open-source models, and the broader implications for safety, security, and global AI competition.

Ernest Hammond III, Associate General Counsel at Meta, shared insights on how the company navigates the balance between openness and responsible safeguards, particularly with its Llama models. Neal Hannan, Senior Director & Associate General Counsel at Databricks, offered a perspective on how Databricks approaches open-source AI from both a business and a regulatory standpoint. Aviya Skowron, Head of Policy at EleutherAI, advocated for more robust transparency and accessibility in AI development, emphasizing a research-driven approach.

From the definition of open-source AI to concerns about safety, regulation, and global competition, the panelists explored the complexities of AI development and its future. Here’s a quick summary of what they had to say.

What Does “Open-Source AI” Really Mean?

Justin opened the discussion by addressing a core question: what does “open-source AI” really mean? While “open-source AI” generally refers to making a model’s weights and the related source code necessary for inference and fine-tuning publicly accessible, the extent of openness can vary. Justin pointed out that some developers use the term “open weights,” while others adhere to a stricter definition of “open-source.” The Open Source Initiative (OSI), for example, defines “open-source AI” as requiring the release of all components necessary to fully recreate the model, including training code and training data.

Aviya explained that EleutherAI’s Pythia model aligns with OSI’s definition of “open-source AI” and that its comprehensive release enables full reproducibility, by allowing others to not only use the model but also verify, modify, and build upon it. Aviya emphasized that EleutherAI’s commitment to transparency and accessibility is central to its mission of fostering open-source AI development, enabling researchers and developers to study, improve, and apply these models without restrictions.

In contrast, some models, such as Meta’s Llama, are released under more restrictive licenses. Ernest emphasized that the company doesn’t follow a “one-size-fits-all” approach, instead implementing Acceptable Use Policies (AUPs) and custom licenses to address risks, particularly for general-purpose models that could be misused. While each release is evaluated on a case-by-case basis, permissive licenses are offered when possible.

The Business and Legal Case for Open-Source AI

The panelists also discussed the business and legal considerations that influence whether companies release AI models as open-source. Justin highlighted that open-source AI presents both significant opportunities and risks. On the one hand, it democratizes access and fosters innovation, but on the other, it can be seen as a competitive liability.

Neal noted that Databricks has deep roots in open-source innovation, dating back to Apache Spark. For Databricks, open-source AI allows customers to fine-tune models with their own data, ensuring bias mitigation, transparency, and customization. Neal explained that while large language models are great for a wide range of tasks, fine-tuning is often necessary to specialize models for specific use cases. This could involve additional training to make the model perform better on a particular document set, address specific questions, or adopt a desired tone or attitude.

Ernest explained that Meta sees open-source releases as a way to drive innovation and engage with the broader ecosystem. While not a core revenue-driver, they accelerate progress by inviting external developers to contribute, as with Llama 2.0, where community efforts expanded its context length, later adopted by Meta. Ernest also highlighted diverse applications of open-source models, from training checkpoints that reveal model evolution to audio generation tools and healthcare translation models aiding doctors in underserved areas.

Neal further noted that Databricks prioritizes helping customers identify and fine-tune models that best suit their needs rather than releasing the largest models. Neal also highlighted how high per-token costs can limit accessibility. This often means choosing smaller, more efficient models over the biggest ones. As demand for transparency grows, some companies are focusing on providing greater visibility into model operations, enabling customers to better leverage AI.

Safety, Security, and the Open-Source Dilemma

One of the most contentious issues in AI policy is whether open-source models pose unique safety risks. Some companies argue that open releases make it harder to prevent misuse, while others believe that transparency outweighs the dangers.

Aviya noted that safety in open-source AI is often misunderstood, with many reducing it to filtering inputs and outputs. Unlike closed models with enforceable terms of service, open-source models lack such controls. Since AUPs often cannot be effectively enforced, EleutherAI does not impose them, instead relying on social pressure within the developer community to discourage misuse.

Meta takes a structured approach to model safety through its Frontier AI Framework, which classifies models based on risk levels. Ernest explained that if a model’s capabilities pose risks like enabling fraud or misinformation, Meta may choose not to release it. The process involves assessing the model’s potential harms, conducting extensive red teaming, and considering mitigations. Ernest also noted that researchers focused on advancing the state of the art may not consider product limitations, making it crucial for risk teams to assess whether a model should be developed and regularly check for biases and other issues.

Neal highlighted Databricks’ focus on AI safety with tools like Llama Guard, observability features, and AI judges. These tools allow users to monitor and mitigate risks in real time and understand safety profiles, ensuring models align with expectations. Justin also referenced a case where an airline’s chatbot provided incorrect pricing information, and the court upheld the decision, highlighting the critical importance of accuracy and safety in AI models.

The International AI Race and Open-Source Models

The panel also discussed the geopolitical implications of open-source AI. Justin pointed out that the U.S.-China tension over AI dominance complicates the landscape. Open-source AI could level the playing field by making powerful models globally accessible, potentially shifting the competition balance.

Neal acknowledged concerns that open-source models like DeepSeek could threaten U.S. economic dominance due to cheaper chips and improved research. However, Neal noted that AI advancements benefit everyone by driving broader adoption, with models potentially becoming exponentially cheaper. Misconceptions about open-source AI were also addressed, with clarification that while DeepSeek’s models are globally accessible, fears about data shifting to China stem from misunderstanding and highlight the need for greater public awareness.

Aviya expressed concern over restrictive AI policies, arguing that AI is fundamentally just math and attempting to control it is futile. Pointing out that export controls won’t address the issue, as math and AI research are universally accessible, Aviya emphasized the importance of collaboration and cautioned against efforts to restrict AI advancements.

Ernest, Aviya, and Neal all agreed that while AI safety is critical, censorship concerns must be taken seriously. Aviya pointed out a major blind spot: copyright disputes. Many AI developers have quietly stopped talking about training on copyrighted data, but litigation has had a chilling effect on research. Aviya warned that ongoing lawsuits and vague transparency laws could stifle innovation in AI.

AI regulation remains a moving target, with governments worldwide debating how much control is too much. The U.S. Department of Commerce’s new export controls (effective January 2025) currently apply only to closed-weight models, but future regulations may expand to open-weight models if deemed too risky. At the recent AI Action Summit in Paris, U.S. Vice President JD Vance warned against over-regulation, stating that AI progress will never come to pass if innovators are deterred from taking risks.

Justin concluded by highlighting that AI policy is still up for grabs, with ongoing debates between deregulation and stricter controls. What some view as essential safety measures, others see as overreach. Justin also referenced the uncertainty surrounding the industry’s future, particularly in relation to transparency laws and regulations like the EU AI Act and California’s AB 2013.

Final Thoughts: What’s Next for Open-Source AI?

Open-source AI remains a hotbed of innovation, but its future depends on how businesses, governments, and developers balance accessibility, competition, and security. The panelists all agreed that openness fuels innovation, but each organization must navigate legal risks, business incentives, and ethical concerns when deciding what to release.

As the panelists discussed these issues, the audience was polled with some revealing questions: Who here has interacted with an LLM-powered chatbot service in the past week? Almost everyone raised their hand. Who has developed software using open-source? Around 50% had. And who has downloaded code and weights of an open-source model themselves? About 25% raised their hands. It’s clear that whether through image generators or chatbots, many of us are using tools powered by open-source models without even realizing it.

With regulations in flux, the only certainty is that AI’s open-source revolution is far from over.