In the latest press release by Meta, Mark Zuckerberg wrote about its open source model and the benefits of it.
In the early days of high-performance computing, major tech companies heavily invested in developing their own closed-source versions of Unix. At the time, it seemed unimaginable that any alternative approach could create such advanced software. However, open-source Linux gradually gained popularity—initially because it allowed developers to modify its code freely and was more affordable. Over time, Linux became more advanced, secure, and supported by a broader ecosystem than any closed Unix system. Today, Linux serves as the industry-standard foundation for both cloud computing and the operating systems running most mobile devices, providing superior products for all.
A similar trajectory is expected for AI. Currently, several tech companies are leading the development of closed AI models. Nevertheless, open-source AI is rapidly closing the gap. Just last year, Llama 2 was only comparable to older generation models. This year, Llama 3 stands toe-to-toe with the most advanced models, even outpacing them in certain aspects. Starting next year, future iterations of Llama will set the industry standard. Even now, Llama excels in terms of openness, modifiability, and cost efficiency.
Today marks significant steps toward making open-source AI the industry norm. The release of Llama 3.1 405B, the first frontier-level open-source AI model, along with new and improved Llama 3.1 70B and 8B models, showcases this commitment. These models offer significantly better cost/performance relative to closed models and provide an excellent foundation for fine-tuning and distilling smaller models.
Beyond releasing these models, collaboration with various companies aims to expand the broader ecosystem. Amazon, Databricks, and NVIDIA are launching comprehensive services to support developers in fine-tuning and distilling their own models. Innovators like Groq have built low-latency, cost-effective inference serving for all new models. These models will be accessible on major cloud platforms, including AWS, Azure, Google, Oracle, and more. Companies like Scale.AI, Dell, and Deloitte are ready to assist enterprises in adopting Llama and training custom models with their data. As the community grows and more companies contribute new services, collectively making Llama the industry standard becomes increasingly tangible.
Meta remains committed to open-source AI. The rationale behind this belief is threefold: it benefits developers, it is advantageous for Meta, and it is ultimately good for the world.
When speaking with developers, CEOs, and government officials worldwide, several themes consistently emerge:
Organizations need to train, fine-tune, and distill their own models. Each organization has unique needs best met with models of different sizes trained or fine-tuned with specific data. On-device tasks and classification tasks require smaller models, while more complex tasks necessitate larger models. With the most advanced Llama models, organizations can continue training with their data and then distill them down to optimally sized models without external oversight.
Organizations need to control their own destiny and avoid being locked into a closed vendor. Many organizations do not want to depend on models they cannot run and control themselves. They do not want closed model providers to have the power to change their model, alter terms of use, or even cease service entirely. They also do not want to be locked into a single cloud that has exclusive rights to a model. Open-source enables a broad ecosystem of companies with compatible toolchains that can be easily navigated.
Organizations need to protect their data. Many handle sensitive data that must be secured and cannot be sent to closed models over cloud APIs. Others simply do not trust closed model providers with their data. Open-source addresses these concerns by allowing models to be run wherever desired. Open-source software tends to be more secure due to its transparent development process.
Organizations need a model that is efficient and affordable to run. Developers can run inference on Llama 3.1 405B on their own infrastructure at roughly 50% of the cost of using closed models like GPT-4o for both user-facing and offline inference tasks.
Many want to invest in an ecosystem that will be the standard for the long term. Open-source advances faster than closed models, and building systems on this architecture offers a long-term advantage.
For Meta, open-source AI aligns with building the best experiences and services for people. Ensuring access to the best technology without being locked into a competitor’s closed ecosystem is crucial. Building services constrained by what other platforms allow has shown that Meta and many other companies could create better services if not restricted by arbitrary rules or monopolistic practices.
Concerns about losing a technical advantage by open-sourcing Llama miss the broader picture for several reasons:
To avoid being restricted by a closed ecosystem over the long term, Llama requires a comprehensive ecosystem of tools, efficiency improvements, silicon optimizations, and integrations. A solitary effort would not suffice.
AI development will remain competitive; thus, open-sourcing any given model does not confer a massive advantage over the next best models at that point in time. The goal is for Llama to remain consistently competitive, efficient, and open across generations.
Meta’s business model is not based on selling access to AI models. Therefore, openly releasing Llama does not undermine revenue or sustainability but rather fosters innovation and collaboration.
Meta has a history of successful open-source projects. Initiatives like Open Compute Project and tools like PyTorch and React have saved billions and spurred innovation. This approach has consistently yielded positive results.
Open-source AI is necessary for a positive AI future. AI has unparalleled potential to boost human productivity, creativity, quality of life, drive economic growth, and advance medical and scientific research. Open-source ensures these benefits are accessible worldwide, preventing power concentration in a few hands and promoting equitable deployment across society.
Safety concerns around open-source AI are valid but evidence suggests that open-source models are safer due to transparency and widespread scrutiny. While intentional harm by bad actors is a concern, the balance of power in a world with widely deployed AI favors security and stability. Larger institutions with robust AI systems can counter threats from less sophisticated actors.
Addressing geopolitical dimensions, democratic nations’ advantage lies in decentralized and open innovation. Some argue for closing models to prevent adversaries from gaining access but this approach is flawed. Espionage and theft are real threats; a closed model landscape would disadvantage startups, universities, and small businesses. A robust open ecosystem supported by leading companies and governments offers the best chance for sustained leadership and innovation.
The future of AI will likely mirror the history of computing shaped by open-source collaboration. The release of Llama 3.1 marks a pivotal moment setting the stage for a new era of AI development. As more developers and partners join this journey making AI accessible and beneficial for all becomes increasingly achievable.