In early August, one day before releasing GPT-5, OpenAI CEO Sam Altman posted an image of the Death Star on social media. It was just the latest declaration by Altman that his new AI model would change the world forever. “We have discovered, invented, whatever you want to call it, something extraordinary that is going to reshape the course of human history,” Altman said in a July interview. He compared his company’s research to the Manhattan Project and said that he felt “useless” compared with OpenAI’s newest invention. Altman, in other words, suggested that GPT-5 would bring society closer to what computer scientists call artificial general intelligence: an AI system that can match or exceed human cognition, including the ability to learn new things.
For years, creating AGI has been the holy grail of many leading AI researchers. Altman and other top technologists, including Anthropic CEO Dario Amodei and computer science professors Yoshua Bengio and Stuart Russell, have been dreaming of constructing superintelligent systems for decades—as well as fearing them. And recently, many of these voices have declared that the day of reckoning is near, telling government officials that whichever country invents AGI first will gain enormous geopolitical advantages. Days before U.S. President Donald Trump’s second inauguration, for example, Altman told Trump that AGI would be achieved within his term—and that Washington needed to prepare.
These declarations have clearly had an effect. Over the last two years, Democratic and Republican politicians alike have been discussing AGI more frequently and exploring policies that could unleash its potential or limit its harms. It is easy to see why. AI is already at the heart of a range of emerging technologies, including robotics, biotechnology, and quantum computing. It is also a central element of U.S.-China competition. AGI could theoretically unlock more (and more impressive) scientific advancements, including the ability to stop others from making similar breakthroughs. In this view, if the United States makes it first, American economic growth might skyrocket and the country could attain an unassailable military advantage.
There is no doubt that AI is a very powerful invention. But when it comes to AGI, the hype has grown out of proportion. Given the limitations of existing systems, it is unlikely that superintelligence is actually imminent, even though AI systems continue to improve. Some prominent computer scientists, such as Andrew Ng, have questioned whether artificial general intelligence will ever be created. For now, and possibly forever, advances in AI are more likely to be iterative, like other general-purpose technologies.
The United States should therefore treat the AI race with China like a marathon, not a sprint. This is especially important given the centrality of AI to Washington’s competition with Beijing. Today, both the country’s new tech firms, like DeepSeek, and existing powerhouses, like Huawei, are increasingly keeping pace with their American counterparts. By emphasizing steady advancements and economic integration, China may now even be ahead of the United States in terms of adopting and using robotics. To win the AI race, Washington thus needs to emphasize practical investments in the development and rapid adoption of AI. It cannot distort U.S. policy by dashing for something that might not exist.
WILDEST DREAMS
In Washington, AGI is a hot topic. In a September 2024 hearing on AI oversight, Connecticut Senator Richard Blumenthal declared that AGI is “here and now—one to three years has been the latest prediction.” In July, South Dakota Senator Mike Rounds introduced a bill requiring the Pentagon to establish an AGI steering committee. The bipartisan U.S.-China Economic and Security Review Commission’s 2024 report argued that AGI demanded a Manhattan Project–level effort to ensure the United States achieved it first. Some officials even believe AGI is about to jeopardize human existence. In June 2025, for instance, Representative Jill Tokuda of Hawaii said that “artificial superintelligence, ASI, is one of the largest existential threats that we face.”
The fixation on AGI goes beyond rhetoric. Former Biden administration officials issued executive orders that regulated AI in part based on concerns that AGI is on the horizon. Trump’s AI Action Plan, released in July, may avoid explicit mentions of AGI. But it emphasizes frontier AI, infrastructure expansions, and an innovation-centric race for technological dominance. It would, in the words of Time magazine, fulfill “many of the greatest policy wishes of the top AI companies—which are all now more certain than ever that AGI is around the corner.”
The argument for dashing toward AGI is simple. An AGI system, the thinking goes, might be able to self-improve simultaneously along multiple dimensions. In doing so, it could quickly surpass what humans are capable of and solve problems that have vexed society for millennia. The company and country that reaches that point first will thus not only achieve enormous financial returns, scientific breakthroughs, and military advancements but also lock out competitors by monopolizing the benefits in ways that restrict the developments of others and that establish the rules of the game. The AI race, then, is really a race to a predetermined, AGI finish line in which the winner not only bursts triumphantly through the ribbon but picks up every trophy and goes home, leaving nothing for even the second- and third-place competitors.
It is unlikely that superintelligence is actually imminent.
Yet there is reason to be skeptical of this framing. For starters, AI researchers can’t even agree on how to define AGI and its capabilities; in other words, no one agrees on where the finish line is. That makes any policy based around achieving it inherently dubious. Instead of a singular creation, AI is more of a broad category of technologies, with many different types of innovations. That means progress is likely to be a complex and ever-changing wave, rather than a straight-line trip.
This is evident in the technology’s most recent developments. Today’s models are making strides in usability. The most advanced large language models, however, still face many of the same challenges they faced in 2022, including shallow reasoning, brittle generalization, a lack of long-term memory, and a lack of genuine metacognition or continual learning—as well, of course, as hallucinations. Since its release, for instance, GPT-5 has looked more like a normal advancement than a transformative breakthrough. As a result, some of AGI’s biggest proponents have started tempering their enthusiasm. At the start of the summer, former Google CEO Eric Schmidt said that AI wasn’t hyped enough; now, he argues that people have become too obsessed with “superintelligent” systems. Similarly, in August, Altman declared that AGI is “not a useful concept.” In some ways, when it comes to AGI, the computer science world may still be where it was in 2002, when the then director of MIT’s AI lab joked that the true definition of AI was “almost implemented.”
Even if some AI models do prove transformative, their effects will be mediated by adoption and diffusion processes—as happens with almost every invention. Consider, for example, electricity. It has generated untold value and utterly transformed the global economy, but it became useful thanks to the thousands of scientists, engineers, inventors, and companies who worked on it over the course of decades. Benjamin Franklin proved lightning was electricity in 1752, Alessandro Volta invented the first battery in 1799, and Nikola Tesla developed alternating current in the late 1880s. Even then, it took many more years before most homes had power outlets. All of these innovations were critical to reaching that eventual endpoint, and no one actor captured the global market for electricity or effectively prevented others from continuing to innovate.
The modern combustion engine provides another case-in-point. It was invented in 1876 by the German engineer Nicholas Otto, but was advanced and improved upon over the course of several decades before automobiles went mainstream. Companies around the world ultimately achieved massive gains from automobiles, not just German ones (although German auto industry is, of course, very successful). Perhaps the most prominent early leader, the Ford Motor Company, was American, and it first dominated the car market thanks to its innovations in production, not engines.
INNOVATION AND ADAPTATION
If AI competition is more likely to span a generation than just a few more years, American officials need to think more about how the country can quickly adopt AI advances and less about how to summon AI’s speculative potential. This is closer to what Beijing does. Although the United States and China are very different and the latter’s approach has its limits, China is moving faster at scaling robots in society, and its AI Plus Initiative emphasizes achieving widespread industry-specific adoption by 2027. The government wants AI to essentially become a part of the country’s infrastructure by 2030. China is also investing in AGI, but Beijing’s emphasis is clearly on quickly scaling, integrating, and applying current and near-term AI capabilities.
To avoid falling behind in AI adoption within the bureaucracy, the United States should launch a large-scale AI literacy initiative across the government. Public employees of all kinds need to know how to use both general AI systems and ones tailored to their jobs. American officials should offer expanded access to AI training both for their particular roles and for general use, including training on issues like automation bias (in which people overestimate the accuracy of AI systems). To do so, Washington can take advantage of the fact that major American companies, including OpenAI and Anthropic, are willing to give public employees and agencies more exposure and access to their technologies, allowing the state, at least for now, to use their large language models virtually for free.
The United States must also modernize its infrastructure and data practices, including within the national security apparatus. Advanced AI models require sophisticated hardware, adequate computing power, and state-of-the-art knowledge management systems to operate effectively. And today, Washington is behind on each. The government has started to make some progress on upgrading its systems, but decades of siloing and bureaucratic processes have created entrenched lags that are hindering innovation. To achieve AI adoption at scale, Washington will likely need to invest billions of dollars in procurement over the next few years, especially for the Pentagon.
Racing toward a myth is not sound policy.
Done right, AI could revolutionize the government’s efficiency. Even if it helps only in mundane areas, such as energy load optimization, cybersecurity and IT, predictive maintenance, logistics, supply chain management, and acquisition paperwork, it will allow larger bureaucracies to overcome or eliminate regulatory hurdles. That could, in turn, fuel more private-sector adoption. Right now, private sector pilot projects with frontier AI sometimes fail to successfully transition from prototype to full capability, often because of integration challenges or misalignment between a proposed AI solution and the problem it targets. By some estimates, more than 80 percent of AI projects fail to deliver results. Industry surveys report that 88 percent of pilots never reach production. The IT company Gartner projects that 40 percent of “agentic AI” deployments—autonomous AI systems capable of planning and executing multi-step tasks with minimal or no human oversight—will be scrapped by 2027. By placing greater value on and demonstrating how AI can be integrated into large, complex bureaucracies, the government can help forge a pathway for private companies, lowering their perceived risks. By adopting AI, Washington can also create demand signal for scalable, near-term AI applications.
But protecting American AI leadership will require the government to do more than just help itself and the private sector. The United States will also need to invest in universities and researchers who can make invaluable technical breakthroughs in AI safety, efficiency, and effectiveness, but lack the capacities of big firms. The Trump administration must therefore follow through on its plan to expand support for the National AI Research Resource, a nascent, government-provided consortium of AI infrastructure that would provide researchers, educators, and students with the specialized tools they need for advanced AI work.
None of these steps means U.S. officials should abandon thinking about AGI. In fact, some of the best policies for ensuring AI leadership today will also hasten the arrival of more advanced systems. Any policy that supports AI research and development, such as the immense investment in technology mandated by the 2022 CHIPS and Science Act, will lead to more sophisticated algorithms. So will continued investment in the country’s power infrastructure, which helps the energy-intensive AI industry grow and function.
But Washington must ensure that the pursuit of AGI does not come at the expense of near-term adoption. Racing toward a myth is not sound policy. Instead, the country’s primary goal must be rapidly scaling practical AI applications—improvements that meet government needs and deliver real efficiencies today and tomorrow. Otherwise, the United States could keep producing the world’s fanciest models. It could lead in algorithm creation. But it will still fall behind countries that make better use of AI innovations.
Loading…
Source link
