The open-source approach to technology development has fueled innovation across the tech industry for decades. Now, it risks fueling adversaries in the geopolitical battles of the 21st century.
People working on open-source projects share their work openly online for others to use, contribute to, and build upon. This approach fosters broad collaboration and speeds innovation, as individuals and companies contribute their time and effort to building toward common goals.
Open-source is a big deal making much of the internet possible, and it fuels innovation. When the world was flat, this drove unprecedented innovation that benefitted all. However, the tech world has developed a new super technology that brings the open-source approach into question.
Artificial intelligence is likely to become the most significant technology of the 21st century. It will transform every business, spawn new products and industries, and forever change the nature of work. Countries that harness it first will yield significant economic and strategic benefits.
In 2017, China announced an ambitious program to become the world’s “major AI innovation center” by 2030. While the United States, Canada and the United Kingdom continue to dominate in AI research, China is highly motivated to close the gap and win the race to develop powerful AI.
Recognizing the importance of maintaining an AI leadership position, the U.S. government has released a series of ever-tightening restrictions on exporting the powerful chips that provide computing capabilities for AI to China. Open-source should be examined closely here, too.
The RISC-V chip architecture is an open-source project that emerged from the University of California, Berkeley, in 2014. RISC-V defines the instruction set architecture of a chip and offers a blueprint for how-to-design computer hardware. RISC-V is used by many Chinese, Indian and Western companies.
China has enthusiastically embraced RISC-V to fuel its domestic chip design ambitions. While RISC-V is focused on traditional central processing unit workloads, new vector instructions added recently by the RISC-V International Foundation accelerate machine learning functions. It’s likely that future extensions will accelerate AI training and inference. One Chinese company has already built a RISC-V-based processor with tensor units, giving it 200 TOPS of impressive AI number-crunching performance.
RISC-V International, the organization that controls the direction of the RISC-V architecture, deliberately relocated itself to Switzerland in 2020 to insulate itself from potential U.S. sanctions and influence. While it’s unlikely this group will develop technology that meaningfully challenges the AI performance of Western companies, they are likely to enable at least some of the future ambitions of a China starved of AI chips by export controls.
Sovereign nations have realized they need their own national AI strategies. They don’t want to rely on a few American companies to provide the essential digital fuel that will power their industries and societies of tomorrow.
The global race to develop AI is on. Which brings us back to open-source.
The three leaders in AI research — OpenAI, Google and Anthropic — create what is known as frontier models, the most powerful models. Some time ago, these companies stopped publishing their research, deciding to take a more responsible approach to releasing AI capabilities. AI services are accessed through a programming interface that allows these companies to restrict access if they determine bad actors are using their models.
The next tier down is more cavalier. Several tech players have released powerful large language models as open-source. These models are typically a generation or so behind the frontier models, but they are still very powerful.
Meta, Stability, Abacus and Databricks — all U.S. companies — are fundamentally committed to open-source, believing that technology as powerful and impactful as artificial intelligence should be democratized rather than restricted to a few giant companies. They have each released powerful large language models for anyone to download and use. The day these models hit the web, AI researchers in China, Iran and other adversaries eagerly download them, dissect them and use them as the foundation for new capabilities.
Open-source technology has fueled widespread innovation in tech for decades. Most people benefit from it hundreds of times a day. However, the continued release of powerful open-source AI models without necessary security controls risks empowering our adversaries and diminishing U.S. leadership.
Would we open-source nuclear know-how? Policymakers should consider the situation and act accordingly.
