Tech Innovation Dies in Darkness
From the Industrial Revolution to the age of AI, it’s clear that openness in technology does far more good than harm
Among the most contentious debates in the budding field of artificial intelligence (AI) policy is the long-term status of so-called open models—AI models whose underlying weights (the set of billions or even trillions of numbers that define the model’s capabilities) are made available for free for anyone to download or modify. Some believe that the release of open AI models is the only way to avoid a catastrophic outcome for humanity. By ensuring that every person, group and country controls its own AI, this line of reasoning goes, we can avoid a scenario where one group monopolizes the power of a single, exceptionally capable model.
The most extreme critics, on the other hand, believe that AI development in general is an existential risk to humanity, and that the release of open AI models is the riskiest approach of them all. Other critics of open models—and some existential risk believers who have pivoted to a more prosaic argument to gain appeal among policymakers—contend that open distribution of models exposes America’s key AI secrets to foreign competitors, most notably China.
All these arguments, however, ignore centuries of technology history. They assume, either explicitly or implicitly, that AI is fundamentally different from prior general-purpose technologies such as electricity, internal combustion, computers and the internet. While it is impossible to predict the trajectory of AI development, history and experience have shown themselves to be a better guide than a priori rationalization time and time again.
If critics of open models believe that history is an ineffective guide for our current challenges, the burden of proof is on them to demonstrate why—a burden they have largely failed to shoulder. This is unfortunate because that history offers clear lessons for technologists and policymakers alike.
Textiles Then
In the 18th century, textile manufacturing was the high-technology consumer good of its day. Britain, the cradle of the First Industrial Revolution, led the way. Its innovations included the spinning jenny, invented in 1764 by James Hargreaves, which allowed a single worker to work on many spools of yarn at once, substantially increasing productivity, and Richard Arkwright’s water frame, the first water-powered textile machine. Taken together, these and other inventions allowed for faster production of stronger and more affordable textiles.
The British government sought to jealously guard these innovations from foreign hands. In 1774, it passed export controls on textile machinery and forbade workers who built such machines from emigrating. Noncompete clauses for those same workers were liberally used and judiciously enforced. This policy climate reinforced a culture of closed innovation: Factory owners worked to secure their factories, seeking to keep out visitors—especially foreign visitors.
Despite this, knowledge, as it always does, managed to escape its bounds. High-skilled British workers, such as Samuel Slater, who was an apprentice of Arkwright, made their way to America and applied British know-how to American industry. Slater partnered with New England financiers to create what some call the first American factory. This financial partnership had several innovative features that presaged the development of American venture capital, as documented by Harvard Business School professor Tom Nicholas in his book “VC: An American History.”
Crafty Americans like Francis Cabot Lowell traveled to British factories, memorizing the designs of their machines and bringing them back to the colony of Massachusetts. The Lowell System created a model for textile manufacturing throughout the United States and turned New England into a domestic hub for this cutting-edge technology. His influence is undeniable: The town of Lowell is named for him and the textile manufacturing powerhouse he built.
Neither Slater nor Lowell was a mere intellectual property thief. Both added their own innovations, including improvements to the efficiency of the technology and to workforce training, learning to integrate technological breakthroughs into the American context. By the mid-19th century, the United States had become a global textile exporter, rivaling if not surpassing Britain’s once-dominant textile sector.
While the precise impact of these policies is difficult to isolate from other economic and political factors, a few facts are clear. First, the British export controls undoubtedly limited the international market for their textile machines, ultimately giving America (and other countries) an opportunity to compete in the global market with their machines as well as the textiles those machines produced. Second, the British policies did not work because economically valuable knowledge is among the hardest things to keep within the walls of a company or the borders of a country.
By the second half of the 19th century, the world had moved on to the Second Industrial Revolution. As the political scientist Jeffrey Ding has pointed out, the United States was not an innovation leader in many of the technologies that defined this era, such as chemical engineering. Researchers at foreign universities—most importantly in Germany—published the most cutting-edge research papers and filed the most innovative patents.
Yet as Ding explains, it was American businesses that were quickest to adapt their industrial processes to suit modern chemical manufacturing. He argues that this was due in large part to close connections between American universities and businesses. The universities were well-positioned to translate the innovations made elsewhere into practical commercial applications. The countries that led in pure innovation, on the other hand, tended to be more siloed and more hesitant to share their innovations with domestic firms.
During this era, the British inventor Henry Bessemer invented a new process for mass-producing steel: the eponymously named Bessemer Process. Like the earlier generation of British textile innovators, Bessemer sought to keep his process a close secret, hoping to monopolize the technology for himself. Americans independently discovered a similar process, causing Bessemer to instigate an arduous legal battle.
Ultimately, the technology would be used most effectively by American industrialists. Ironically, Bessemer’s legacy lives on in the American venture capital giant Bessemer Venture Partners—founded not by Bessemer himself but by Henry Phipps Jr., an American industrialist who built his fortune on Bessemer’s innovation.
Lessons Learned for Tech Today
From just this brief look into the history of textile technology development, we can glean a few important lessons that we’d be smart to apply to the development of AI today. The first of these lessons is that technological development looks more like the gradual accumulation of sedimentary layers than it does the impact of a meteor. New technologies, and general-purpose technologies in particular, tend to be most effective when they are integrated with existing technologies rather than standing on their own. While popular historical narratives about technology tend to focus on singular innovators like Thomas Edison and Steve Jobs, much of the benefit of new technologies is derived from discovering how to integrate those innovations into practical life—a process often called technology diffusion.
Second, the benefits of open innovation usually far exceed the costs. Openness quickens the pace of innovation, allowing for the cross-pollination of ideas between researchers and engineers. It also accrues soft power to the countries whose firms and inventors lead in innovation. Successful technologists—those whose products lead in both innovation and mass adoption—often set the hard and soft standards for the use of their inventions. And because information technologies such as AI are embedded with cultural, political and philosophical values, the countries whose innovations lead the world are also exporting these values to billions of people.
Finally, openness greatly aids the process of diffusion because effective diffusion often requires flexibility and extensibility from new technologies—classic features of open and competitive technology marketplaces. For these reasons, countries that attempt to lock down their technological secrets often harm themselves more than their competitors.
For an American, the examples of British hoarding of technology secrets present a troubling parallel to our country’s technology policies today. It is hard not to compare Britain’s aggressive measures to keep emerging technology out of the hands of a rising—and hungrier—global competitor with our own efforts to restrict Chinese access to AI computing chips and semiconductor manufacturing equipment. Perhaps these specific measures are appropriate. But the broad sweep of history suggests that export controls, particularly on AI models themselves, are a losing recipe to maintaining our current leadership status in the field, and may even backfire in unpredictable ways.
The Open-Vs.-Closed Debate in AI
Open-source is a decades-old distribution model for software. It is perhaps the best contemporary example of the benefits openness can deliver to both companies and countries. Open-source software has humble roots. The academics and other tinkerers developing the earliest forms of modern software realized that it was easier to share and collaborate than it was to build every piece of software they needed by themselves.
Gradually, this became a political, sometimes almost religious, movement. Famously, Richard Stallman, the creator of the license that still governs the release of much open-source software (licenses play a key role in all software, including open-source), said that open-source was about freedom “as in speech, not as in beer”—though it was free in the beer sense as well.
But what began as an outgrowth of 1960s West Coast counterculture has morphed into the digital lifeblood of the modern economy. The fundamental needs of early computing pioneers remained the same even for large corporations, particularly those without software expertise. It is easier and faster to begin one’s software engineering efforts on a shared foundation than it is to build everything from scratch. Indeed, given the volume of software used in modern industry, it is scarcely possible to imagine a world without such a shared, open foundation. And because open-source software can be shared with and scrutinized by anyone in the world, keeping it secure and up-to-date is also a global collective effort, with engineers from trillion-dollar firms contributing alongside solo coders in their proverbial basements.
This model has paid benefits that would have been unimaginable to many even as recently as the 1990s, when open-source software was already in widespread use. Open-source software is at the heart of almost every modern smartphone, in the form of the Unix operating system kernel for Apple’s iOS and the Linux operating system for Google’s Android. Nearly every website is powered by a kaleidoscopic variety of open-source software tools, from the Apache webserver to the MySQL database. Most of the programming languages used to write modern software are themselves open-source.
There is scarcely a modern good—digital or physical—one can identify that was not somehow enabled by open-source software, because inasmuch as computers were involved in making that good, so too was open-source software. It should come as no surprise that one of the only Western internet platforms not censored by the Chinese government is Microsoft’s GitHub, the dominant repository of open-source software.
Open AI models are a continuation of this powerful tradition. Models of this variety can be further divided into two categories: “open-weight” models, where the model developer only makes the weights available publicly, and fully open-source models, whose weights, associated code and training data are released publicly.
The potential benefits of open-source AI models are similar to those of open-source software in general. Many companies in the broader economy, curious about adopting AI in their business processes, demand the flexibility and ownership uniquely enabled by the open-source model. Some of them are also reluctant (or legally unable) to share their proprietary corporate data with closed-model developers, again necessitating the use of an open model.
Academic research and other efforts to advance AI safety, along with our understanding of how large AI models work, also often require the transparency of an open model, which allows researchers to “look inside” a model and probe its inner workings in a way that is impossible with closed-source models.
As mentioned earlier, critics of open AI models allege that they pose grave dangers, either to humanity itself or to the United States in particular. A report by the AI safety company Gladstone, commissioned by the U.S. State Department, recommended making open-source AI development a criminal offense.
Numerous AI safety and policy nonprofits, such as the Center for AI Safety or the Center for AI Policy, have proposed regulations that would make open-source AI development effectively impossible, if not criminalize it. SB 1047, a bill introduced in the California legislature by state Sen. Scott Wiener and written in close collaboration with the Center for AI Safety, has been criticized as making the most powerful AI models challenging or impossible to release as open-source.
Despite the heated rhetoric and ominous policy signals, American firms continue to develop some of the best open large language models in the world. While Google, Apple, Microsoft and many others have released open-weight and open-source models, Meta stands out as having grounded its AI strategy in open releases. Its Llama 3 model sits near the top of the open models on the LMSYS Chatbot Arena Leaderboard, a website that ranks generative AI models based on user feedback (Google’s Gemma model was the top-ranked open model as of early July 2024).
The Leaderboard’s top 10 slots, however, are filled almost entirely by closed models from OpenAI, Anthropic and Google. (Yi, a closed model from the Chinese company 01.AI, occupies ninth place.)
Under the surface, however, Chinese companies and academic researchers continue to publish open models and research results that move the global field forward. Open models from Alibaba and the startup DeepSeek, for example, are close behind the top American open models and have surpassed the performance of earlier versions of OpenAI’s GPT-4. DeepSeek’s models in particular stand out. The company has gained a positive reputation in the global AI community for several excellent models and research papers.
Another Chinese company, Zhipu AI, has raised eyebrows for the license it attaches to its open models, which requires any company that uses the model for commercial ends to register with it and mandates that any legal disputes relating to the license or the model be adjudicated in Chinese courts. This is a small taste of what could happen if the United States forfeits its lead in open AI development. Ultimately, the leaders in a technological field set the standards for global use of their inventions, be they legal, cultural or political.
And of course, because language models in particular have political and philosophical values embedded deep within them, it is easy to imagine what other losses America might incur if it abandons open AI models. The alarm that some American elites felt when they saw how TikTok systematically de-emphasized pro-Israel content on the platform in the wake of the October 7 attacks by Hamas and ensuing war in Gaza will be a mere preview of what might happen if Chinese language models (even ones that speak English) dominate the global AI field.
Building a Smart AI Policy Framework
The AI policy conversation is in a curious place today. Nearly every U.S. policy and lobbying nonprofit with the word “AI” in its name supports policies that pose an existential threat to open-source AI development. Indeed, most of these groups were formed because of fears that AI represents an existential risk to humanity—a concern that, thus far, has little empirical evidence to support it.
Yet at the same time, individuals and groups as diverse as the Heritage Foundation, the Center for American Progress, the Cato Institute and Federal Trade Commissioner Lina Khan all have voiced support for continued open AI development. This is likely because open-source is deeply resonant with long-held American ideals. It is software’s version of the First Amendment or the Enlightenment Republic of Letters. It is hard to wrest such values from a country’s intellectual genome. In a world replete with threats to liberal democracy, we should seek to preserve those values wherever we can.
Currently, the United States is the leader in both open and closed AI development. It can maintain this position through continued investment, ingenuity and hard work. It can lose it by choice, including the choice of policymakers with good intentions. The United States may already have gone too far with export controls and mandatory secrecy. But at the very least, applying export controls to AI models—rather than the enabling hardware—could be a ruinous move, not least because export controls make open-source releases virtually impossible.
More broadly, the culture of secrecy that has developed around AI development in the United States could be a long-term handicap. While many Chinese firms (and those of other countries) publish leading-edge research publicly, in the United States that research is increasingly cloistered inside the frontier AI companies: Google DeepMind, Anthropic and OpenAI. Only Meta stands out among that group for continuing to publish its research. If anything, then, policymakers should be looking for ways to nudge AI companies toward open release of models and research rather than away from it.
Francois Chollet, an AI researcher at Google, recently said in a podcast interview with Dwarkesh Patel that OpenAI’s attempts to close off its AI development efforts “basically set back progress” in the field of AI “by quite a few years, probably like 5-10 … they caused this complete closing down of frontier research publishing.” Even if he was exaggerating, the historical parallels explored earlier should give us pause: Innovation does not tend to flourish in darkness.